1 /*
2 * Copyright (c) 1997, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/block.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/convertnode.hpp"
35 #include "opto/idealGraphPrinter.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/machnode.hpp"
38 #include "opto/opcodes.hpp"
39 #include "opto/phaseX.hpp"
40 #include "opto/regalloc.hpp"
41 #include "opto/rootnode.hpp"
42 #include "utilities/macros.hpp"
43 #include "utilities/powerOfTwo.hpp"
44
45 //=============================================================================
46 #define NODE_HASH_MINIMUM_SIZE 255
47
48 //------------------------------NodeHash---------------------------------------
49 NodeHash::NodeHash(Arena *arena, uint est_max_size) :
50 _a(arena),
51 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
52 _inserts(0), _insert_limit( insert_limit() ),
53 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) )
54 #ifndef PRODUCT
55 , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0),
56 _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0),
57 _total_inserts(0), _total_insert_probes(0)
58 #endif
59 {
60 // _sentinel must be in the current node space
61 _sentinel = new ProjNode(nullptr, TypeFunc::Control);
62 memset(_table,0,sizeof(Node*)*_max);
63 }
64
65 //------------------------------hash_find--------------------------------------
66 // Find in hash table
67 Node *NodeHash::hash_find( const Node *n ) {
68 // ((Node*)n)->set_hash( n->hash() );
69 uint hash = n->hash();
70 if (hash == Node::NO_HASH) {
71 NOT_PRODUCT( _lookup_misses++ );
72 return nullptr;
73 }
74 uint key = hash & (_max-1);
75 uint stride = key | 0x01;
76 NOT_PRODUCT( _look_probes++ );
77 Node *k = _table[key]; // Get hashed value
78 if( !k ) { // ?Miss?
79 NOT_PRODUCT( _lookup_misses++ );
80 return nullptr; // Miss!
81 }
82
83 int op = n->Opcode();
84 uint req = n->req();
85 while( 1 ) { // While probing hash table
86 if( k->req() == req && // Same count of inputs
87 k->Opcode() == op ) { // Same Opcode
88 for( uint i=0; i<req; i++ )
89 if( n->in(i)!=k->in(i)) // Different inputs?
90 goto collision; // "goto" is a speed hack...
91 if( n->cmp(*k) ) { // Check for any special bits
92 NOT_PRODUCT( _lookup_hits++ );
93 return k; // Hit!
94 }
95 }
96 collision:
97 NOT_PRODUCT( _look_probes++ );
98 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
99 k = _table[key]; // Get hashed value
100 if( !k ) { // ?Miss?
101 NOT_PRODUCT( _lookup_misses++ );
102 return nullptr; // Miss!
103 }
104 }
105 ShouldNotReachHere();
106 return nullptr;
107 }
108
109 //------------------------------hash_find_insert-------------------------------
110 // Find in hash table, insert if not already present
111 // Used to preserve unique entries in hash table
112 Node *NodeHash::hash_find_insert( Node *n ) {
113 // n->set_hash( );
114 uint hash = n->hash();
115 if (hash == Node::NO_HASH) {
116 NOT_PRODUCT( _lookup_misses++ );
117 return nullptr;
118 }
119 uint key = hash & (_max-1);
120 uint stride = key | 0x01; // stride must be relatively prime to table siz
121 uint first_sentinel = 0; // replace a sentinel if seen.
122 NOT_PRODUCT( _look_probes++ );
123 Node *k = _table[key]; // Get hashed value
124 if( !k ) { // ?Miss?
125 NOT_PRODUCT( _lookup_misses++ );
126 _table[key] = n; // Insert into table!
127 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
128 check_grow(); // Grow table if insert hit limit
129 return nullptr; // Miss!
130 }
131 else if( k == _sentinel ) {
132 first_sentinel = key; // Can insert here
133 }
134
135 int op = n->Opcode();
136 uint req = n->req();
137 while( 1 ) { // While probing hash table
138 if( k->req() == req && // Same count of inputs
139 k->Opcode() == op ) { // Same Opcode
140 for( uint i=0; i<req; i++ )
141 if( n->in(i)!=k->in(i)) // Different inputs?
142 goto collision; // "goto" is a speed hack...
143 if( n->cmp(*k) ) { // Check for any special bits
144 NOT_PRODUCT( _lookup_hits++ );
145 return k; // Hit!
146 }
147 }
148 collision:
149 NOT_PRODUCT( _look_probes++ );
150 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
151 k = _table[key]; // Get hashed value
152 if( !k ) { // ?Miss?
153 NOT_PRODUCT( _lookup_misses++ );
154 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
155 _table[key] = n; // Insert into table!
156 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
157 check_grow(); // Grow table if insert hit limit
158 return nullptr; // Miss!
159 }
160 else if( first_sentinel == 0 && k == _sentinel ) {
161 first_sentinel = key; // Can insert here
162 }
163
164 }
165 ShouldNotReachHere();
166 return nullptr;
167 }
168
169 //------------------------------hash_insert------------------------------------
170 // Insert into hash table
171 void NodeHash::hash_insert( Node *n ) {
172 // // "conflict" comments -- print nodes that conflict
173 // bool conflict = false;
174 // n->set_hash();
175 uint hash = n->hash();
176 if (hash == Node::NO_HASH) {
177 return;
178 }
179 check_grow();
180 uint key = hash & (_max-1);
181 uint stride = key | 0x01;
182
183 while( 1 ) { // While probing hash table
184 NOT_PRODUCT( _insert_probes++ );
185 Node *k = _table[key]; // Get hashed value
186 if( !k || (k == _sentinel) ) break; // Found a slot
187 assert( k != n, "already inserted" );
188 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; }
189 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
190 }
191 _table[key] = n; // Insert into table!
192 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
193 // if( conflict ) { n->dump(); }
194 }
195
196 //------------------------------hash_delete------------------------------------
197 // Replace in hash table with sentinel
198 bool NodeHash::hash_delete( const Node *n ) {
199 Node *k;
200 uint hash = n->hash();
201 if (hash == Node::NO_HASH) {
202 NOT_PRODUCT( _delete_misses++ );
203 return false;
204 }
205 uint key = hash & (_max-1);
206 uint stride = key | 0x01;
207 DEBUG_ONLY( uint counter = 0; );
208 for( ; /* (k != nullptr) && (k != _sentinel) */; ) {
209 DEBUG_ONLY( counter++ );
210 NOT_PRODUCT( _delete_probes++ );
211 k = _table[key]; // Get hashed value
212 if( !k ) { // Miss?
213 NOT_PRODUCT( _delete_misses++ );
214 return false; // Miss! Not in chain
215 }
216 else if( n == k ) {
217 NOT_PRODUCT( _delete_hits++ );
218 _table[key] = _sentinel; // Hit! Label as deleted entry
219 DEBUG_ONLY(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
220 return true;
221 }
222 else {
223 // collision: move through table with prime offset
224 key = (key + stride/*7*/) & (_max-1);
225 assert( counter <= _insert_limit, "Cycle in hash-table");
226 }
227 }
228 ShouldNotReachHere();
229 return false;
230 }
231
232 //------------------------------round_up---------------------------------------
233 // Round up to nearest power of 2
234 uint NodeHash::round_up(uint x) {
235 x += (x >> 2); // Add 25% slop
236 return MAX2(16U, round_up_power_of_2(x));
237 }
238
239 //------------------------------grow-------------------------------------------
240 // Grow _table to next power of 2 and insert old entries
241 void NodeHash::grow() {
242 // Record old state
243 uint old_max = _max;
244 Node **old_table = _table;
245 // Construct new table with twice the space
246 #ifndef PRODUCT
247 _grows++;
248 _total_inserts += _inserts;
249 _total_insert_probes += _insert_probes;
250 _insert_probes = 0;
251 #endif
252 _inserts = 0;
253 _max = _max << 1;
254 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
255 memset(_table,0,sizeof(Node*)*_max);
256 _insert_limit = insert_limit();
257 // Insert old entries into the new table
258 for( uint i = 0; i < old_max; i++ ) {
259 Node *m = *old_table++;
260 if( !m || m == _sentinel ) continue;
261 DEBUG_ONLY(m->exit_hash_lock()); // Unlock the node upon removal from old table.
262 hash_insert(m);
263 }
264 }
265
266 //------------------------------clear------------------------------------------
267 // Clear all entries in _table to null but keep storage
268 void NodeHash::clear() {
269 #ifdef ASSERT
270 // Unlock all nodes upon removal from table.
271 for (uint i = 0; i < _max; i++) {
272 Node* n = _table[i];
273 if (!n || n == _sentinel) continue;
274 n->exit_hash_lock();
275 }
276 #endif
277
278 memset( _table, 0, _max * sizeof(Node*) );
279 }
280
281 //-----------------------remove_useless_nodes----------------------------------
282 // Remove useless nodes from value table,
283 // implementation does not depend on hash function
284 void NodeHash::remove_useless_nodes(VectorSet &useful) {
285
286 // Dead nodes in the hash table inherited from GVN should not replace
287 // existing nodes, remove dead nodes.
288 uint max = size();
289 Node *sentinel_node = sentinel();
290 for( uint i = 0; i < max; ++i ) {
291 Node *n = at(i);
292 if(n != nullptr && n != sentinel_node && !useful.test(n->_idx)) {
293 DEBUG_ONLY(n->exit_hash_lock()); // Unlock the node when removed
294 _table[i] = sentinel_node; // Replace with placeholder
295 }
296 }
297 }
298
299
300 void NodeHash::check_no_speculative_types() {
301 #ifdef ASSERT
302 uint max = size();
303 Unique_Node_List live_nodes;
304 Compile::current()->identify_useful_nodes(live_nodes);
305 Node *sentinel_node = sentinel();
306 for (uint i = 0; i < max; ++i) {
307 Node *n = at(i);
308 if (n != nullptr &&
309 n != sentinel_node &&
310 n->is_Type() &&
311 live_nodes.member(n)) {
312 TypeNode* tn = n->as_Type();
313 const Type* t = tn->type();
314 const Type* t_no_spec = t->remove_speculative();
315 assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup");
316 }
317 }
318 #endif
319 }
320
321 #ifndef PRODUCT
322 //------------------------------dump-------------------------------------------
323 // Dump statistics for the hash table
324 void NodeHash::dump() {
325 _total_inserts += _inserts;
326 _total_insert_probes += _insert_probes;
327 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
328 if (WizardMode) {
329 for (uint i=0; i<_max; i++) {
330 if (_table[i])
331 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
332 }
333 }
334 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max);
335 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
336 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
337 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
338 // sentinels increase lookup cost, but not insert cost
339 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
340 assert( _inserts+(_inserts>>3) < _max, "table too full" );
341 assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
342 }
343 }
344
345 Node *NodeHash::find_index(uint idx) { // For debugging
346 // Find an entry by its index value
347 for( uint i = 0; i < _max; i++ ) {
348 Node *m = _table[i];
349 if( !m || m == _sentinel ) continue;
350 if( m->_idx == (uint)idx ) return m;
351 }
352 return nullptr;
353 }
354 #endif
355
356 #ifdef ASSERT
357 NodeHash::~NodeHash() {
358 // Unlock all nodes upon destruction of table.
359 if (_table != (Node**)badAddress) clear();
360 }
361 #endif
362
363 // Add users of 'n' that match 'predicate' to worklist
364 template <class Predicate>
365 static void add_users_to_worklist_if(Unique_Node_List& worklist, const Node* n, Predicate predicate) {
366 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
367 Node* u = n->fast_out(i);
368 if (predicate(u)) {
369 worklist.push(u);
370 }
371 }
372 }
373
374 //=============================================================================
375 //------------------------------PhaseRemoveUseless-----------------------------
376 // 1) Use a breadthfirst walk to collect useful nodes reachable from root.
377 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN* gvn, Unique_Node_List& worklist, PhaseNumber phase_num) : Phase(phase_num) {
378 C->print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
379 // Implementation requires an edge from root to each SafePointNode
380 // at a backward branch. Inserted in add_safepoint().
381
382 // Identify nodes that are reachable from below, useful.
383 C->identify_useful_nodes(_useful);
384 // Update dead node list
385 C->update_dead_node_list(_useful);
386
387 // Remove all useless nodes from PhaseValues' recorded types
388 // Must be done before disconnecting nodes to preserve hash-table-invariant
389 gvn->remove_useless_nodes(_useful.member_set());
390
391 // Remove all useless nodes from future worklist
392 worklist.remove_useless_nodes(_useful.member_set());
393
394 // Disconnect 'useless' nodes that are adjacent to useful nodes
395 C->disconnect_useless_nodes(_useful, worklist);
396 }
397
398 //=============================================================================
399 //------------------------------PhaseRenumberLive------------------------------
400 // First, remove useless nodes (equivalent to identifying live nodes).
401 // Then, renumber live nodes.
402 //
403 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure.
404 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the
405 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique
406 // value in the range [0, x).
407 //
408 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is
409 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes).
410 //
411 // The PhaseRenumberLive phase updates two data structures with the new node IDs.
412 // (1) The "worklist" is "C->igvn_worklist()", which is to collect which nodes need to
413 // be processed by IGVN after removal of the useless nodes.
414 // (2) Type information "gvn->types()" (same as "C->types()") maps every node ID to
415 // the node's type. The mapping is updated to use the new node IDs as well. We
416 // create a new map, and swap it with the old one.
417 //
418 // Other data structures used by the compiler are not updated. The hash table for value
419 // numbering ("C->node_hash()", referenced by PhaseValue::_table) is not updated because
420 // computing the hash values is not based on node IDs.
421 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn,
422 Unique_Node_List& worklist,
423 PhaseNumber phase_num) :
424 PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live),
425 _new_type_array(C->comp_arena()),
426 _old2new_map(C->unique(), C->unique(), -1),
427 _is_pass_finished(false),
428 _live_node_count(C->live_nodes())
429 {
430 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place");
431 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes");
432 assert(_delayed.size() == 0, "should be empty");
433 assert(&worklist == C->igvn_worklist(), "reference still same as the one from Compile");
434 assert(&gvn->types() == C->types(), "reference still same as that from Compile");
435
436 GrowableArray<Node_Notes*>* old_node_note_array = C->node_note_array();
437 if (old_node_note_array != nullptr) {
438 int new_size = (_useful.size() >> 8) + 1; // The node note array uses blocks, see C->_log2_node_notes_block_size
439 new_size = MAX2(8, new_size);
440 C->set_node_note_array(new (C->comp_arena()) GrowableArray<Node_Notes*> (C->comp_arena(), new_size, 0, nullptr));
441 C->grow_node_notes(C->node_note_array(), new_size);
442 }
443
444 assert(worklist.is_subset_of(_useful), "only useful nodes should still be in the worklist");
445
446 // Iterate over the set of live nodes.
447 for (uint current_idx = 0; current_idx < _useful.size(); current_idx++) {
448 Node* n = _useful.at(current_idx);
449
450 const Type* type = gvn->type_or_null(n);
451 _new_type_array.map(current_idx, type);
452
453 assert(_old2new_map.at(n->_idx) == -1, "already seen");
454 _old2new_map.at_put(n->_idx, current_idx);
455
456 if (old_node_note_array != nullptr) {
457 Node_Notes* nn = C->locate_node_notes(old_node_note_array, n->_idx);
458 C->set_node_notes_at(current_idx, nn);
459 }
460
461 n->set_idx(current_idx); // Update node ID.
462
463 if (update_embedded_ids(n) < 0) {
464 _delayed.push(n); // has embedded IDs; handle later
465 }
466 }
467
468 // VectorSet in Unique_Node_Set must be recomputed, since IDs have changed.
469 worklist.recompute_idx_set();
470
471 assert(_live_node_count == _useful.size(), "all live nodes must be processed");
472
473 _is_pass_finished = true; // pass finished; safe to process delayed updates
474
475 while (_delayed.size() > 0) {
476 Node* n = _delayed.pop();
477 int no_of_updates = update_embedded_ids(n);
478 assert(no_of_updates > 0, "should be updated");
479 }
480
481 // Replace the compiler's type information with the updated type information.
482 gvn->types().swap(_new_type_array);
483
484 // Update the unique node count of the compilation to the number of currently live nodes.
485 C->set_unique(_live_node_count);
486
487 // Set the dead node count to 0 and reset dead node list.
488 C->reset_dead_node_list();
489 }
490
491 int PhaseRenumberLive::new_index(int old_idx) {
492 assert(_is_pass_finished, "not finished");
493 if (_old2new_map.at(old_idx) == -1) { // absent
494 // Allocate a placeholder to preserve uniqueness
495 _old2new_map.at_put(old_idx, _live_node_count);
496 _live_node_count++;
497 }
498 return _old2new_map.at(old_idx);
499 }
500
501 int PhaseRenumberLive::update_embedded_ids(Node* n) {
502 int no_of_updates = 0;
503 if (n->is_Phi()) {
504 PhiNode* phi = n->as_Phi();
505 if (phi->_inst_id != -1) {
506 if (!_is_pass_finished) {
507 return -1; // delay
508 }
509 int new_idx = new_index(phi->_inst_id);
510 assert(new_idx != -1, "");
511 phi->_inst_id = new_idx;
512 no_of_updates++;
513 }
514 if (phi->_inst_mem_id != -1) {
515 if (!_is_pass_finished) {
516 return -1; // delay
517 }
518 int new_idx = new_index(phi->_inst_mem_id);
519 assert(new_idx != -1, "");
520 phi->_inst_mem_id = new_idx;
521 no_of_updates++;
522 }
523 }
524
525 const Type* type = _new_type_array.fast_lookup(n->_idx);
526 if (type != nullptr && type->isa_oopptr() && type->is_oopptr()->is_known_instance()) {
527 if (!_is_pass_finished) {
528 return -1; // delay
529 }
530 int old_idx = type->is_oopptr()->instance_id();
531 int new_idx = new_index(old_idx);
532 const Type* new_type = type->is_oopptr()->with_instance_id(new_idx);
533 _new_type_array.map(n->_idx, new_type);
534 no_of_updates++;
535 }
536
537 return no_of_updates;
538 }
539
540 void PhaseValues::init_con_caches() {
541 memset(_icons,0,sizeof(_icons));
542 memset(_lcons,0,sizeof(_lcons));
543 memset(_zcons,0,sizeof(_zcons));
544 }
545
546 PhaseIterGVN* PhaseValues::is_IterGVN() {
547 return (_phase == PhaseValuesType::iter_gvn || _phase == PhaseValuesType::ccp) ? static_cast<PhaseIterGVN*>(this) : nullptr;
548 }
549
550 //--------------------------------find_int_type--------------------------------
551 const TypeInt* PhaseValues::find_int_type(Node* n) {
552 if (n == nullptr) return nullptr;
553 // Call type_or_null(n) to determine node's type since we might be in
554 // parse phase and call n->Value() may return wrong type.
555 // (For example, a phi node at the beginning of loop parsing is not ready.)
556 const Type* t = type_or_null(n);
557 if (t == nullptr) return nullptr;
558 return t->isa_int();
559 }
560
561
562 //-------------------------------find_long_type--------------------------------
563 const TypeLong* PhaseValues::find_long_type(Node* n) {
564 if (n == nullptr) return nullptr;
565 // (See comment above on type_or_null.)
566 const Type* t = type_or_null(n);
567 if (t == nullptr) return nullptr;
568 return t->isa_long();
569 }
570
571 //------------------------------~PhaseValues-----------------------------------
572 #ifndef PRODUCT
573 PhaseValues::~PhaseValues() {
574 // Statistics for NodeHash
575 _table.dump();
576 // Statistics for value progress and efficiency
577 if( PrintCompilation && Verbose && WizardMode ) {
578 tty->print("\n%sValues: %d nodes ---> " UINT64_FORMAT "/%d (%d)",
579 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values());
580 if( made_transforms() != 0 ) {
581 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() );
582 } else {
583 tty->cr();
584 }
585 }
586 }
587 #endif
588
589 //------------------------------makecon----------------------------------------
590 ConNode* PhaseValues::makecon(const Type* t) {
591 assert(t->singleton(), "must be a constant");
592 assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
593 switch (t->base()) { // fast paths
594 case Type::Half:
595 case Type::Top: return (ConNode*) C->top();
596 case Type::Int: return intcon( t->is_int()->get_con() );
597 case Type::Long: return longcon( t->is_long()->get_con() );
598 default: break;
599 }
600 if (t->is_zero_type())
601 return zerocon(t->basic_type());
602 return uncached_makecon(t);
603 }
604
605 //--------------------------uncached_makecon-----------------------------------
606 // Make an idealized constant - one of ConINode, ConPNode, etc.
607 ConNode* PhaseValues::uncached_makecon(const Type *t) {
608 assert(t->singleton(), "must be a constant");
609 ConNode* x = ConNode::make(t);
610 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
611 if (k == nullptr) {
612 set_type(x, t); // Missed, provide type mapping
613 GrowableArray<Node_Notes*>* nna = C->node_note_array();
614 if (nna != nullptr) {
615 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
616 loc->clear(); // do not put debug info on constants
617 }
618 } else {
619 x->destruct(this); // Hit, destroy duplicate constant
620 x = k; // use existing constant
621 }
622 return x;
623 }
624
625 //------------------------------intcon-----------------------------------------
626 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))"
627 ConINode* PhaseValues::intcon(jint i) {
628 // Small integer? Check cache! Check that cached node is not dead
629 if (i >= _icon_min && i <= _icon_max) {
630 ConINode* icon = _icons[i-_icon_min];
631 if (icon != nullptr && icon->in(TypeFunc::Control) != nullptr)
632 return icon;
633 }
634 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
635 assert(icon->is_Con(), "");
636 if (i >= _icon_min && i <= _icon_max)
637 _icons[i-_icon_min] = icon; // Cache small integers
638 return icon;
639 }
640
641 //------------------------------longcon----------------------------------------
642 // Fast long constant.
643 ConLNode* PhaseValues::longcon(jlong l) {
644 // Small integer? Check cache! Check that cached node is not dead
645 if (l >= _lcon_min && l <= _lcon_max) {
646 ConLNode* lcon = _lcons[l-_lcon_min];
647 if (lcon != nullptr && lcon->in(TypeFunc::Control) != nullptr)
648 return lcon;
649 }
650 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
651 assert(lcon->is_Con(), "");
652 if (l >= _lcon_min && l <= _lcon_max)
653 _lcons[l-_lcon_min] = lcon; // Cache small integers
654 return lcon;
655 }
656 ConNode* PhaseValues::integercon(jlong l, BasicType bt) {
657 if (bt == T_INT) {
658 return intcon(checked_cast<jint>(l));
659 }
660 assert(bt == T_LONG, "not an integer");
661 return longcon(l);
662 }
663
664
665 //------------------------------zerocon-----------------------------------------
666 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
667 ConNode* PhaseValues::zerocon(BasicType bt) {
668 assert((uint)bt <= _zcon_max, "domain check");
669 ConNode* zcon = _zcons[bt];
670 if (zcon != nullptr && zcon->in(TypeFunc::Control) != nullptr)
671 return zcon;
672 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
673 _zcons[bt] = zcon;
674 return zcon;
675 }
676
677
678
679 //=============================================================================
680 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) {
681 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape);
682 if (i == nullptr) {
683 i = k->Ideal(this, can_reshape);
684 }
685 return i;
686 }
687
688 //------------------------------transform--------------------------------------
689 // Return a node which computes the same function as this node, but
690 // in a faster or cheaper fashion.
691 Node* PhaseGVN::transform(Node* n) {
692 NOT_PRODUCT( set_transforms(); )
693
694 // Apply the Ideal call in a loop until it no longer applies
695 Node* k = n;
696 Node* i = apply_ideal(k, /*can_reshape=*/false);
697 NOT_PRODUCT(uint loop_count = 1;)
698 while (i != nullptr) {
699 assert(i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
700 k = i;
701 #ifdef ASSERT
702 if (loop_count >= K + C->live_nodes()) {
703 dump_infinite_loop_info(i, "PhaseGVN::transform");
704 }
705 #endif
706 i = apply_ideal(k, /*can_reshape=*/false);
707 NOT_PRODUCT(loop_count++;)
708 }
709 NOT_PRODUCT(if (loop_count != 0) { set_progress(); })
710
711 // If brand new node, make space in type array.
712 ensure_type_or_null(k);
713
714 // Since I just called 'Value' to compute the set of run-time values
715 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
716 // cache Value. Later requests for the local phase->type of this Node can
717 // use the cached Value instead of suffering with 'bottom_type'.
718 const Type* t = k->Value(this); // Get runtime Value set
719 assert(t != nullptr, "value sanity");
720 if (type_or_null(k) != t) {
721 #ifndef PRODUCT
722 // Do not count initial visit to node as a transformation
723 if (type_or_null(k) == nullptr) {
724 inc_new_values();
725 set_progress();
726 }
727 #endif
728 set_type(k, t);
729 // If k is a TypeNode, capture any more-precise type permanently into Node
730 k->raise_bottom_type(t);
731 }
732
733 if (t->singleton() && !k->is_Con()) {
734 set_progress();
735 return makecon(t); // Turn into a constant
736 }
737
738 // Now check for Identities
739 i = k->Identity(this); // Look for a nearby replacement
740 if (i != k) { // Found? Return replacement!
741 set_progress();
742 return i;
743 }
744
745 // Global Value Numbering
746 i = hash_find_insert(k); // Insert if new
747 if (i && (i != k)) {
748 // Return the pre-existing node
749 set_progress();
750 return i;
751 }
752
753 // Return Idealized original
754 return k;
755 }
756
757 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) {
758 if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) {
759 return false;
760 }
761 if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) {
762 return false;
763 }
764 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
765 int i = 0;
766 while (d != n) {
767 n = IfNode::up_one_dom(n, linear_only);
768 i++;
769 if (n == nullptr || i >= 100) {
770 return false;
771 }
772 }
773 return true;
774 }
775
776 #ifdef ASSERT
777 //------------------------------dead_loop_check--------------------------------
778 // Check for a simple dead loop when a data node references itself directly
779 // or through an other data node excluding cons and phis.
780 void PhaseGVN::dead_loop_check(Node* n) {
781 // Phi may reference itself in a loop.
782 if (n == nullptr || n->is_dead_loop_safe() || n->is_CFG()) {
783 return;
784 }
785
786 // Do 2 levels check and only data inputs.
787 for (uint i = 1; i < n->req(); i++) {
788 Node* in = n->in(i);
789 if (in == n) {
790 n->dump_bfs(100, nullptr, "");
791 fatal("Dead loop detected, node references itself: %s (%d)",
792 n->Name(), n->_idx);
793 }
794
795 if (in == nullptr || in->is_dead_loop_safe()) {
796 continue;
797 }
798 for (uint j = 1; j < in->req(); j++) {
799 if (in->in(j) == n) {
800 n->dump_bfs(100, nullptr, "");
801 fatal("Dead loop detected, node input references current node: %s (%d) -> %s (%d)",
802 in->Name(), in->_idx, n->Name(), n->_idx);
803 }
804 if (in->in(j) == in) {
805 n->dump_bfs(100, nullptr, "");
806 fatal("Dead loop detected, node input references itself: %s (%d)",
807 in->Name(), in->_idx);
808 }
809 }
810 }
811 }
812
813
814 /**
815 * Dumps information that can help to debug the problem. A debug
816 * build fails with an assert.
817 */
818 void PhaseGVN::dump_infinite_loop_info(Node* n, const char* where) {
819 n->dump(4);
820 assert(false, "infinite loop in %s", where);
821 }
822 #endif
823
824 //=============================================================================
825 //------------------------------PhaseIterGVN-----------------------------------
826 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
827 PhaseIterGVN::PhaseIterGVN(PhaseIterGVN* igvn) : _delay_transform(igvn->_delay_transform),
828 _worklist(*C->igvn_worklist())
829 {
830 _phase = PhaseValuesType::iter_gvn;
831 assert(&_worklist == &igvn->_worklist, "sanity");
832 }
833
834 //------------------------------PhaseIterGVN-----------------------------------
835 // Initialize from scratch
836 PhaseIterGVN::PhaseIterGVN() : _delay_transform(false),
837 _worklist(*C->igvn_worklist())
838 {
839 _phase = PhaseValuesType::iter_gvn;
840 uint max;
841
842 // Dead nodes in the hash table inherited from GVN were not treated as
843 // roots during def-use info creation; hence they represent an invisible
844 // use. Clear them out.
845 max = _table.size();
846 for( uint i = 0; i < max; ++i ) {
847 Node *n = _table.at(i);
848 if(n != nullptr && n != _table.sentinel() && n->outcnt() == 0) {
849 if( n->is_top() ) continue;
850 // If remove_useless_nodes() has run, we expect no such nodes left.
851 assert(false, "remove_useless_nodes missed this node");
852 hash_delete(n);
853 }
854 }
855
856 // Any Phis or Regions on the worklist probably had uses that could not
857 // make more progress because the uses were made while the Phis and Regions
858 // were in half-built states. Put all uses of Phis and Regions on worklist.
859 max = _worklist.size();
860 for( uint j = 0; j < max; j++ ) {
861 Node *n = _worklist.at(j);
862 uint uop = n->Opcode();
863 if( uop == Op_Phi || uop == Op_Region ||
864 n->is_Type() ||
865 n->is_Mem() )
866 add_users_to_worklist(n);
867 }
868 }
869
870 void PhaseIterGVN::shuffle_worklist() {
871 if (_worklist.size() < 2) return;
872 for (uint i = _worklist.size() - 1; i >= 1; i--) {
873 uint j = C->random() % (i + 1);
874 swap(_worklist.adr()[i], _worklist.adr()[j]);
875 }
876 }
877
878 #ifndef PRODUCT
879 void PhaseIterGVN::verify_step(Node* n) {
880 if (is_verify_def_use()) {
881 ResourceMark rm;
882 VectorSet visited;
883 Node_List worklist;
884
885 _verify_window[_verify_counter % _verify_window_size] = n;
886 ++_verify_counter;
887 if (C->unique() < 1000 || 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
888 ++_verify_full_passes;
889 worklist.push(C->root());
890 Node::verify(-1, visited, worklist);
891 return;
892 }
893 for (int i = 0; i < _verify_window_size; i++) {
894 Node* n = _verify_window[i];
895 if (n == nullptr) {
896 continue;
897 }
898 if (n->in(0) == NodeSentinel) { // xform_idom
899 _verify_window[i] = n->in(1);
900 --i;
901 continue;
902 }
903 // Typical fanout is 1-2, so this call visits about 6 nodes.
904 if (!visited.test_set(n->_idx)) {
905 worklist.push(n);
906 }
907 }
908 Node::verify(4, visited, worklist);
909 }
910 }
911
912 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) {
913 const Type* newtype = type_or_null(n);
914 if (nn != n || oldtype != newtype) {
915 C->print_method(PHASE_AFTER_ITER_GVN_STEP, 5, n);
916 }
917 if (TraceIterativeGVN) {
918 uint wlsize = _worklist.size();
919 if (nn != n) {
920 // print old node
921 tty->print("< ");
922 if (oldtype != newtype && oldtype != nullptr) {
923 oldtype->dump();
924 }
925 do { tty->print("\t"); } while (tty->position() < 16);
926 tty->print("<");
927 n->dump();
928 }
929 if (oldtype != newtype || nn != n) {
930 // print new node and/or new type
931 if (oldtype == nullptr) {
932 tty->print("* ");
933 } else if (nn != n) {
934 tty->print("> ");
935 } else {
936 tty->print("= ");
937 }
938 if (newtype == nullptr) {
939 tty->print("null");
940 } else {
941 newtype->dump();
942 }
943 do { tty->print("\t"); } while (tty->position() < 16);
944 nn->dump();
945 }
946 if (Verbose && wlsize < _worklist.size()) {
947 tty->print(" Push {");
948 while (wlsize != _worklist.size()) {
949 Node* pushed = _worklist.at(wlsize++);
950 tty->print(" %d", pushed->_idx);
951 }
952 tty->print_cr(" }");
953 }
954 if (nn != n) {
955 // ignore n, it might be subsumed
956 verify_step((Node*) nullptr);
957 }
958 }
959 }
960
961 void PhaseIterGVN::init_verifyPhaseIterGVN() {
962 _verify_counter = 0;
963 _verify_full_passes = 0;
964 for (int i = 0; i < _verify_window_size; i++) {
965 _verify_window[i] = nullptr;
966 }
967 #ifdef ASSERT
968 // Verify that all modified nodes are on _worklist
969 Unique_Node_List* modified_list = C->modified_nodes();
970 while (modified_list != nullptr && modified_list->size()) {
971 Node* n = modified_list->pop();
972 if (!n->is_Con() && !_worklist.member(n)) {
973 n->dump();
974 fatal("modified node is not on IGVN._worklist");
975 }
976 }
977 #endif
978 }
979
980 void PhaseIterGVN::verify_PhaseIterGVN(bool deep_revisit_converged) {
981 #ifdef ASSERT
982 // Verify nodes with changed inputs.
983 Unique_Node_List* modified_list = C->modified_nodes();
984 while (modified_list != nullptr && modified_list->size()) {
985 Node* n = modified_list->pop();
986 if (!n->is_Con()) { // skip Con nodes
987 n->dump();
988 fatal("modified node was not processed by IGVN.transform_old()");
989 }
990 }
991 #endif
992
993 C->verify_graph_edges();
994 if (is_verify_def_use() && PrintOpto) {
995 if (_verify_counter == _verify_full_passes) {
996 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
997 (int) _verify_full_passes);
998 } else {
999 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
1000 (int) _verify_counter, (int) _verify_full_passes);
1001 }
1002 }
1003
1004 #ifdef ASSERT
1005 if (modified_list != nullptr) {
1006 while (modified_list->size() > 0) {
1007 Node* n = modified_list->pop();
1008 n->dump();
1009 assert(false, "VerifyIterativeGVN: new modified node was added");
1010 }
1011 }
1012
1013 verify_optimize(deep_revisit_converged);
1014 #endif
1015 }
1016 #endif /* PRODUCT */
1017
1018 #ifdef ASSERT
1019 /**
1020 * Dumps information that can help to debug the problem. A debug
1021 * build fails with an assert.
1022 */
1023 void PhaseIterGVN::dump_infinite_loop_info(Node* n, const char* where) {
1024 n->dump(4);
1025 _worklist.dump();
1026 assert(false, "infinite loop in %s", where);
1027 }
1028
1029 /**
1030 * Prints out information about IGVN if the 'verbose' option is used.
1031 */
1032 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) {
1033 if (TraceIterativeGVN && Verbose) {
1034 tty->print(" Pop ");
1035 n->dump();
1036 if ((num_processed % 100) == 0) {
1037 _worklist.print_set();
1038 }
1039 }
1040 }
1041 #endif /* ASSERT */
1042
1043 bool PhaseIterGVN::needs_deep_revisit(const Node* n) const {
1044 // LoadNode::Value() -> can_see_stored_value() walks up through many memory
1045 // nodes. LoadNode::Ideal() -> find_previous_store() also walks up to 50
1046 // nodes through stores and arraycopy nodes.
1047 if (n->is_Load()) {
1048 return true;
1049 }
1050 // CmpPNode::sub() -> detect_ptr_independence() -> all_controls_dominate()
1051 // walks CFG dominator relationships extensively. This only triggers when
1052 // both inputs are oop pointers (subnode.cpp:984).
1053 if (n->Opcode() == Op_CmpP) {
1054 const Type* t1 = type_or_null(n->in(1));
1055 const Type* t2 = type_or_null(n->in(2));
1056 return t1 != nullptr && t1->isa_oopptr() &&
1057 t2 != nullptr && t2->isa_oopptr();
1058 }
1059 // IfNode::Ideal() -> search_identical() walks up the CFG dominator tree.
1060 // RangeCheckNode::Ideal() scans up to ~999 nodes up the chain.
1061 // CountedLoopEndNode/LongCountedLoopEndNode::Ideal() via simple_subsuming
1062 // looks for dominating test that subsumes the current test.
1063 switch (n->Opcode()) {
1064 case Op_If:
1065 case Op_RangeCheck:
1066 case Op_CountedLoopEnd:
1067 case Op_LongCountedLoopEnd:
1068 return true;
1069 default:
1070 break;
1071 }
1072 return false;
1073 }
1074
1075 bool PhaseIterGVN::drain_worklist() {
1076 uint loop_count = 1;
1077 const int max_live_nodes_increase_per_iteration = NodeLimitFudgeFactor * 5;
1078 while (_worklist.size() != 0) {
1079 if (C->check_node_count(max_live_nodes_increase_per_iteration, "Out of nodes")) {
1080 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1081 return true;
1082 }
1083 Node* n = _worklist.pop();
1084 if (loop_count >= K * C->live_nodes()) {
1085 DEBUG_ONLY(dump_infinite_loop_info(n, "PhaseIterGVN::drain_worklist");)
1086 C->record_method_not_compilable("infinite loop in PhaseIterGVN::drain_worklist");
1087 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1088 return true;
1089 }
1090 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, _num_processed++);)
1091 if (n->outcnt() != 0) {
1092 NOT_PRODUCT(const Type* oldtype = type_or_null(n));
1093 // Do the transformation
1094 DEBUG_ONLY(int live_nodes_before = C->live_nodes();)
1095 Node* nn = transform_old(n);
1096 DEBUG_ONLY(int live_nodes_after = C->live_nodes();)
1097 // Ensure we did not increase the live node count with more than
1098 // max_live_nodes_increase_per_iteration during the call to transform_old.
1099 DEBUG_ONLY(int increase = live_nodes_after - live_nodes_before;)
1100 assert(increase < max_live_nodes_increase_per_iteration,
1101 "excessive live node increase in single iteration of IGVN: %d "
1102 "(should be at most %d)",
1103 increase, max_live_nodes_increase_per_iteration);
1104 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);)
1105 } else if (!n->is_top()) {
1106 remove_dead_node(n, NodeOrigin::Graph);
1107 }
1108 loop_count++;
1109 }
1110 return false;
1111 }
1112
1113 void PhaseIterGVN::push_deep_revisit_candidates() {
1114 ResourceMark rm;
1115 Unique_Node_List all_nodes;
1116 all_nodes.push(C->root());
1117 for (uint j = 0; j < all_nodes.size(); j++) {
1118 Node* n = all_nodes.at(j);
1119 if (needs_deep_revisit(n)) {
1120 _worklist.push(n);
1121 }
1122 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1123 all_nodes.push(n->fast_out(i));
1124 }
1125 }
1126 }
1127
1128 bool PhaseIterGVN::deep_revisit() {
1129 // Re-process nodes that inspect the graph deeply. After the main worklist drains, walk
1130 // the graph to find all live deep-inspection nodes and push them to the worklist
1131 // for re-evaluation. If any produce changes, drain the worklist again.
1132 // Repeat until stable. This mirrors PhaseCCP::analyze()'s revisit loop.
1133 const uint max_deep_revisit_rounds = 10; // typically converges in <2 rounds
1134 uint round = 0;
1135 for (; round < max_deep_revisit_rounds; round++) {
1136 push_deep_revisit_candidates();
1137 if (_worklist.size() == 0) {
1138 break; // No deep-inspection nodes to revisit, done.
1139 }
1140
1141 #ifndef PRODUCT
1142 uint candidates = _worklist.size();
1143 uint n_if = 0; uint n_rc = 0; uint n_load = 0; uint n_cmpp = 0; uint n_cle = 0; uint n_lcle = 0;
1144 if (TraceIterativeGVN) {
1145 for (uint i = 0; i < _worklist.size(); i++) {
1146 Node* n = _worklist.at(i);
1147 switch (n->Opcode()) {
1148 case Op_If: n_if++; break;
1149 case Op_RangeCheck: n_rc++; break;
1150 case Op_CountedLoopEnd: n_cle++; break;
1151 case Op_LongCountedLoopEnd: n_lcle++; break;
1152 case Op_CmpP: n_cmpp++; break;
1153 default: if (n->is_Load()) n_load++; break;
1154 }
1155 }
1156 }
1157 #endif
1158
1159 // Convergence: if the drain does not make progress (no Ideal, Value, Identity or GVN changes),
1160 // we are at a fixed point. We use made_progress() rather than live_nodes because live_nodes
1161 // misses non-structural changes like a LoadNode dropping its control input.
1162 uint progress_before = made_progress();
1163 if (drain_worklist()) {
1164 return false;
1165 }
1166 uint progress = made_progress() - progress_before;
1167
1168 #ifndef PRODUCT
1169 if (TraceIterativeGVN) {
1170 tty->print("deep_revisit round %u: %u candidates (If=%u RC=%u Load=%u CmpP=%u CLE=%u LCLE=%u), progress=%u (%s)",
1171 round, candidates, n_if, n_rc, n_load, n_cmpp, n_cle, n_lcle, progress, progress != 0 ? "changed" : "converged");
1172 if (C->method() != nullptr) {
1173 tty->print(", ");
1174 C->method()->print_short_name(tty);
1175 }
1176 tty->cr();
1177 }
1178 #endif
1179
1180 if (progress == 0) {
1181 break;
1182 }
1183 }
1184 return round < max_deep_revisit_rounds;
1185 }
1186
1187 void PhaseIterGVN::optimize(bool deep) {
1188 bool deep_revisit_converged = false;
1189 DEBUG_ONLY(_num_processed = 0;)
1190 NOT_PRODUCT(init_verifyPhaseIterGVN();)
1191 NOT_PRODUCT(C->reset_igv_phase_iter(PHASE_AFTER_ITER_GVN_STEP);)
1192 C->print_method(PHASE_BEFORE_ITER_GVN, 3);
1193 if (StressIGVN) {
1194 shuffle_worklist();
1195 }
1196
1197 // Pull from worklist and transform the node.
1198 if (drain_worklist()) {
1199 return;
1200 }
1201
1202 if (deep && UseDeepIGVNRevisit) {
1203 deep_revisit_converged = deep_revisit();
1204 if (C->failing()) {
1205 return;
1206 }
1207 }
1208
1209 NOT_PRODUCT(verify_PhaseIterGVN(deep_revisit_converged);)
1210 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1211 }
1212
1213 #ifdef ASSERT
1214 void PhaseIterGVN::verify_optimize(bool deep_revisit_converged) {
1215 assert(_worklist.size() == 0, "igvn worklist must be empty before verify");
1216
1217 if (is_verify_Value() ||
1218 is_verify_Ideal() ||
1219 is_verify_Identity() ||
1220 is_verify_invariants()) {
1221 ResourceMark rm;
1222 Unique_Node_List worklist;
1223 // BFS all nodes, starting at root
1224 worklist.push(C->root());
1225 for (uint j = 0; j < worklist.size(); ++j) {
1226 Node* n = worklist.at(j);
1227 // If we get an assert here, check why the reported node was not processed again in IGVN.
1228 // We should either make sure that this node is properly added back to the IGVN worklist
1229 // in PhaseIterGVN::add_users_to_worklist to update it again or add an exception
1230 // in the verification methods below if that is not possible for some reason (like Load nodes).
1231 if (is_verify_Value()) {
1232 verify_Value_for(n, deep_revisit_converged /* strict */);
1233 }
1234 if (is_verify_Ideal()) {
1235 verify_Ideal_for(n, false /* can_reshape */, deep_revisit_converged);
1236 verify_Ideal_for(n, true /* can_reshape */, deep_revisit_converged);
1237 }
1238 if (is_verify_Identity()) {
1239 verify_Identity_for(n);
1240 }
1241 if (is_verify_invariants()) {
1242 verify_node_invariants_for(n);
1243 }
1244
1245 // traverse all inputs and outputs
1246 for (uint i = 0; i < n->req(); i++) {
1247 if (n->in(i) != nullptr) {
1248 worklist.push(n->in(i));
1249 }
1250 }
1251 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1252 worklist.push(n->fast_out(i));
1253 }
1254 }
1255 }
1256
1257 verify_empty_worklist(nullptr);
1258 }
1259
1260 void PhaseIterGVN::verify_empty_worklist(Node* node) {
1261 // Verify that the igvn worklist is empty. If no optimization happened, then
1262 // nothing needs to be on the worklist.
1263 if (_worklist.size() == 0) { return; }
1264
1265 stringStream ss; // Print as a block without tty lock.
1266 for (uint j = 0; j < _worklist.size(); j++) {
1267 Node* n = _worklist.at(j);
1268 ss.print("igvn.worklist[%d] ", j);
1269 n->dump("\n", false, &ss);
1270 }
1271 if (_worklist.size() != 0 && node != nullptr) {
1272 ss.print_cr("Previously optimized:");
1273 node->dump("\n", false, &ss);
1274 }
1275 tty->print_cr("%s", ss.as_string());
1276 assert(false, "igvn worklist must still be empty after verify");
1277 }
1278
1279 // Check that type(n) == n->Value(), asserts if we have a failure.
1280 // We have a list of exceptions, see detailed comments in code.
1281 // (1) Integer "widen" changes, but the range is the same.
1282 // (2) LoadNode performs deep traversals. Load is not notified for changes far away.
1283 // (3) CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1284 void PhaseIterGVN::verify_Value_for(const Node* n, bool strict) {
1285 // If we assert inside type(n), because the type is still a null, then maybe
1286 // the node never went through gvn.transform, which would be a bug.
1287 const Type* told = type(n);
1288 const Type* tnew = n->Value(this);
1289 if (told == tnew) {
1290 return;
1291 }
1292 // Exception (1)
1293 // Integer "widen" changes, but range is the same.
1294 if (told->isa_integer(tnew->basic_type()) != nullptr) { // both either int or long
1295 const TypeInteger* t0 = told->is_integer(tnew->basic_type());
1296 const TypeInteger* t1 = tnew->is_integer(tnew->basic_type());
1297 if (t0->lo_as_long() == t1->lo_as_long() &&
1298 t0->hi_as_long() == t1->hi_as_long()) {
1299 return; // ignore integer widen
1300 }
1301 }
1302 // Exception (2)
1303 // LoadNode performs deep traversals. Load is not notified for changes far away.
1304 if (!strict && n->is_Load() && !told->singleton()) {
1305 // MemNode::can_see_stored_value looks up through many memory nodes,
1306 // which means we would need to notify modifications from far up in
1307 // the inputs all the way down to the LoadNode. We don't do that.
1308 return;
1309 }
1310 // Exception (3)
1311 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1312 if (!strict && n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
1313 // SubNode::Value
1314 // CmpPNode::sub
1315 // MemNode::detect_ptr_independence
1316 // MemNode::all_controls_dominate
1317 // We find all controls of a pointer load, and see if they dominate the control of
1318 // an allocation. If they all dominate, we know the allocation is after (independent)
1319 // of the pointer load, and we can say the pointers are different. For this we call
1320 // n->dominates(sub, nlist) to check if controls n of the pointer load dominate the
1321 // control sub of the allocation. The problems is that sometimes dominates answers
1322 // false conservatively, and later it can determine that it is indeed true. Loops with
1323 // Region heads can lead to giving up, whereas LoopNodes can be skipped easier, and
1324 // so the traversal becomes more powerful. This is difficult to remedy, we would have
1325 // to notify the CmpP of CFG updates. Luckily, we recompute CmpP::Value during CCP
1326 // after loop-opts, so that should take care of many of these cases.
1327 return;
1328 }
1329
1330 stringStream ss; // Print as a block without tty lock.
1331 ss.cr();
1332 ss.print_cr("Missed Value optimization:");
1333 n->dump_bfs(3, nullptr, "", &ss);
1334 ss.print_cr("Current type:");
1335 told->dump_on(&ss);
1336 ss.cr();
1337 ss.print_cr("Optimized type:");
1338 tnew->dump_on(&ss);
1339 ss.cr();
1340 tty->print_cr("%s", ss.as_string());
1341
1342 switch (_phase) {
1343 case PhaseValuesType::iter_gvn:
1344 assert(false, "Missed Value optimization opportunity in PhaseIterGVN for %s",n->Name());
1345 break;
1346 case PhaseValuesType::ccp:
1347 assert(false, "PhaseCCP not at fixpoint: analysis result may be unsound for %s", n->Name());
1348 break;
1349 default:
1350 assert(false, "Unexpected phase");
1351 break;
1352 }
1353 }
1354
1355 // Check that all Ideal optimizations that could be done were done.
1356 // Asserts if it found missed optimization opportunities or encountered unexpected changes, and
1357 // returns normally otherwise (no missed optimization, or skipped verification).
1358 void PhaseIterGVN::verify_Ideal_for(Node* n, bool can_reshape, bool deep_revisit_converged) {
1359 if (!deep_revisit_converged && needs_deep_revisit(n)) {
1360 return;
1361 }
1362
1363 // First, we check a list of exceptions, where we skip verification,
1364 // because there are known cases where Ideal can optimize after IGVN.
1365 // Some may be expected and cannot be fixed, and others should be fixed.
1366 switch (n->Opcode()) {
1367 // RegionNode::Ideal does "Skip around the useless IF diamond".
1368 // 245 IfTrue === 244
1369 // 258 If === 245 257
1370 // 259 IfTrue === 258 [[ 263 ]]
1371 // 260 IfFalse === 258 [[ 263 ]]
1372 // 263 Region === 263 260 259 [[ 263 268 ]]
1373 // to
1374 // 245 IfTrue === 244
1375 // 263 Region === 263 245 _ [[ 263 268 ]]
1376 //
1377 // "Useless" means that there is no code in either branch of the If.
1378 // I found a case where this was not done yet during IGVN.
1379 // Why does the Region not get added to IGVN worklist when the If diamond becomes useless?
1380 //
1381 // Found with:
1382 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1383 case Op_Region:
1384 return;
1385
1386 // In AddNode::Ideal, we call "commute", which swaps the inputs so
1387 // that smaller idx are first. Tracking it back, it led me to
1388 // PhaseIdealLoop::remix_address_expressions which swapped the edges.
1389 //
1390 // Example:
1391 // Before PhaseIdealLoop::remix_address_expressions
1392 // 154 AddI === _ 12 144
1393 // After PhaseIdealLoop::remix_address_expressions
1394 // 154 AddI === _ 144 12
1395 // After AddNode::Ideal
1396 // 154 AddI === _ 12 144
1397 //
1398 // I suspect that the node should be added to the IGVN worklist after
1399 // PhaseIdealLoop::remix_address_expressions
1400 //
1401 // This is the only case I looked at, there may be others. Found like this:
1402 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1403 //
1404 // The following hit the same logic in PhaseIdealLoop::remix_address_expressions.
1405 //
1406 // Note: currently all of these fail also for other reasons, for example
1407 // because of "commute" doing the reordering with the phi below. Once
1408 // that is resolved, we can come back to this issue here.
1409 //
1410 // case Op_AddD:
1411 // case Op_AddI:
1412 // case Op_AddL:
1413 // case Op_AddF:
1414 // case Op_MulI:
1415 // case Op_MulL:
1416 // case Op_MulF:
1417 // case Op_MulD:
1418 // if (n->in(1)->_idx > n->in(2)->_idx) {
1419 // // Expect "commute" to revert this case.
1420 // return false;
1421 // }
1422 // break; // keep verifying
1423
1424 // AddFNode::Ideal calls "commute", which can reorder the inputs for this:
1425 // Check for tight loop increments: Loop-phi of Add of loop-phi
1426 // It wants to take the phi into in(1):
1427 // 471 Phi === 435 38 390
1428 // 390 AddF === _ 471 391
1429 //
1430 // Other Associative operators are also affected equally.
1431 //
1432 // Investigate why this does not happen earlier during IGVN.
1433 //
1434 // Found with:
1435 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1436 // -XX:VerifyIterativeGVN=1110
1437 case Op_AddD:
1438 //case Op_AddI: // Also affected for other reasons, see case further down.
1439 //case Op_AddL: // Also affected for other reasons, see case further down.
1440 case Op_AddF:
1441 case Op_MulI:
1442 case Op_MulL:
1443 case Op_MulF:
1444 case Op_MulD:
1445 case Op_MinF:
1446 case Op_MinD:
1447 case Op_MaxF:
1448 case Op_MaxD:
1449 // XorINode::Ideal
1450 // Found with:
1451 // compiler/intrinsics/chacha/TestChaCha20.java
1452 // -XX:VerifyIterativeGVN=1110
1453 case Op_XorI:
1454 case Op_XorL:
1455 // It seems we may have similar issues with the HF cases.
1456 // Found with aarch64:
1457 // compiler/vectorization/TestFloat16VectorOperations.java
1458 // -XX:VerifyIterativeGVN=1110
1459 case Op_AddHF:
1460 case Op_MulHF:
1461 case Op_MaxHF:
1462 case Op_MinHF:
1463 return;
1464
1465 // In MulNode::Ideal the edges can be swapped to help value numbering:
1466 //
1467 // // We are OK if right is a constant, or right is a load and
1468 // // left is a non-constant.
1469 // if( !(t2->singleton() ||
1470 // (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) {
1471 // if( t1->singleton() || // Left input is a constant?
1472 // // Otherwise, sort inputs (commutativity) to help value numbering.
1473 // (in(1)->_idx > in(2)->_idx) ) {
1474 // swap_edges(1, 2);
1475 //
1476 // Why was this not done earlier during IGVN?
1477 //
1478 // Found with:
1479 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1480 // -XX:VerifyIterativeGVN=1110
1481 case Op_AndI:
1482 // Same for AndL.
1483 // Found with:
1484 // compiler/intrinsics/bigInteger/MontgomeryMultiplyTest.java
1485 // -XX:VerifyIterativeGVN=1110
1486 case Op_AndL:
1487 return;
1488
1489 // SubLNode::Ideal does transform like:
1490 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1491 //
1492 // In IGVN before verification:
1493 // 8423 ConvI2L === _ 3519 [[ 8424 ]] #long:-2
1494 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1495 // 8424 AddL === _ 8422 8423 [[ 8383 ]] !orig=[8382]
1496 // 8016 ConL === 0 [[ 8383 ]] #long:0
1497 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1498 //
1499 // And then in verification:
1500 // 8338 ConL === 0 [[ 8339 8424 ]] #long:-2 <----- Was constant folded.
1501 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1502 // 8424 AddL === _ 8422 8338 [[ 8383 ]] !orig=[8382]
1503 // 8016 ConL === 0 [[ 8383 ]] #long:0
1504 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1505 //
1506 // So the form changed from:
1507 // c1 - (y + [8423 ConvI2L])
1508 // to
1509 // c1 - (y + -2)
1510 // but the SubL was not added to the IGVN worklist. Investigate why.
1511 // There could be other issues too.
1512 //
1513 // There seems to be a related AddL IGVN optimization that triggers
1514 // the same SubL optimization, so investigate that too.
1515 //
1516 // Found with:
1517 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1518 case Op_SubL:
1519 return;
1520
1521 // SubINode::Ideal does
1522 // Convert "x - (y+c0)" into "(x-y) - c0" AND
1523 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1524 //
1525 // Investigate why this does not yet happen during IGVN.
1526 //
1527 // Found with:
1528 // test/hotspot/jtreg/compiler/c2/IVTest.java
1529 // -XX:VerifyIterativeGVN=1110
1530 case Op_SubI:
1531 return;
1532
1533 // AddNode::IdealIL does transform like:
1534 // Convert x + (con - y) into "(x - y) + con"
1535 //
1536 // In IGVN before verification:
1537 // 8382 ConvI2L
1538 // 8381 ConvI2L === _ 791 [[ 8383 ]] #long:0
1539 // 8383 SubL === _ 8381 8382
1540 // 8168 ConvI2L
1541 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1542 //
1543 // And then in verification:
1544 // 8424 AddL
1545 // 8016 ConL === 0 [[ 8383 ]] #long:0 <--- Was constant folded.
1546 // 8383 SubL === _ 8016 8424
1547 // 8168 ConvI2L
1548 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1549 //
1550 // So the form changed from:
1551 // x + (ConvI2L(0) - [8382 ConvI2L])
1552 // to
1553 // x + (0 - [8424 AddL])
1554 // but the AddL was not added to the IGVN worklist. Investigate why.
1555 // There could be other issues, too. For example with "commute", see above.
1556 //
1557 // Found with:
1558 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1559 case Op_AddL:
1560 return;
1561
1562 // SubTypeCheckNode::Ideal calls SubTypeCheckNode::verify_helper, which does
1563 // Node* cmp = phase->transform(new CmpPNode(subklass, in(SuperKlass)));
1564 // record_for_cleanup(cmp, phase);
1565 // This verification code in the Ideal code creates new nodes, and checks
1566 // if they fold in unexpected ways. This means some nodes are created and
1567 // added to the worklist, even if the SubTypeCheck is not optimized. This
1568 // goes agains the assumption of the verification here, which assumes that
1569 // if the node is not optimized, then no new nodes should be created, and
1570 // also no nodes should be added to the worklist.
1571 // I see two options:
1572 // 1) forbid what verify_helper does, because for each Ideal call it
1573 // uses memory and that is suboptimal. But it is not clear how that
1574 // verification can be done otherwise.
1575 // 2) Special case the verification here. Probably the new nodes that
1576 // were just created are dead, i.e. they are not connected down to
1577 // root. We could verify that, and remove those nodes from the graph
1578 // by setting all their inputs to nullptr. And of course we would
1579 // have to remove those nodes from the worklist.
1580 // Maybe there are other options too, I did not dig much deeper yet.
1581 //
1582 // Found with:
1583 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1584 case Op_SubTypeCheck:
1585 return;
1586
1587 // LoopLimitNode::Ideal when stride is constant power-of-2, we can do a lowering
1588 // to other nodes: Conv, Add, Sub, Mul, And ...
1589 //
1590 // 107 ConI === 0 [[ ... ]] #int:2
1591 // 84 LoadRange === _ 7 83
1592 // 50 ConI === 0 [[ ... ]] #int:0
1593 // 549 LoopLimit === _ 50 84 107
1594 //
1595 // I stepped backward, to see how the node was generated, and I found that it was
1596 // created in PhaseIdealLoop::exact_limit and not changed since. It is added to the
1597 // IGVN worklist. I quickly checked when it goes into LoopLimitNode::Ideal after
1598 // that, and it seems we want to skip lowering it until after loop-opts, but never
1599 // add call record_for_post_loop_opts_igvn. This would be an easy fix, but there
1600 // could be other issues too.
1601 //
1602 // Fond with:
1603 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1604 case Op_LoopLimit:
1605 return;
1606
1607 // PhiNode::Ideal calls split_flow_path, which tries to do this:
1608 // "This optimization tries to find two or more inputs of phi with the same constant
1609 // value. It then splits them into a separate Phi, and according Region."
1610 //
1611 // Example:
1612 // 130 DecodeN === _ 129
1613 // 50 ConP === 0 [[ 18 91 99 18 ]] #null
1614 // 18 Phi === 14 50 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1615 //
1616 // turns into:
1617 //
1618 // 50 ConP === 0 [[ 99 91 18 ]] #null
1619 // 130 DecodeN === _ 129 [[ 18 ]]
1620 // 18 Phi === 14 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1621 //
1622 // We would have to investigate why this optimization does not happen during IGVN.
1623 // There could also be other issues - I did not investigate further yet.
1624 //
1625 // Found with:
1626 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1627 case Op_Phi:
1628 return;
1629
1630 // MemBarNode::Ideal does "Eliminate volatile MemBars for scalar replaced objects".
1631 // For examle "The allocated object does not escape".
1632 //
1633 // It seems the difference to earlier calls to MemBarNode::Ideal, is that there
1634 // alloc->as_Allocate()->does_not_escape_thread() returned false, but in verification
1635 // it returned true. Why does the MemBarStoreStore not get added to the IGVN
1636 // worklist when this change happens?
1637 //
1638 // Found with:
1639 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1640 case Op_MemBarStoreStore:
1641 return;
1642
1643 // ConvI2LNode::Ideal converts
1644 // 648 AddI === _ 583 645 [[ 661 ]]
1645 // 661 ConvI2L === _ 648 [[ 664 ]] #long:0..maxint-1:www
1646 // into
1647 // 772 ConvI2L === _ 645 [[ 773 ]] #long:-120..maxint-61:www
1648 // 771 ConvI2L === _ 583 [[ 773 ]] #long:60..120:www
1649 // 773 AddL === _ 771 772 [[ ]]
1650 //
1651 // We have to investigate why this does not happen during IGVN in this case.
1652 // There could also be other issues - I did not investigate further yet.
1653 //
1654 // Found with:
1655 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1656 case Op_ConvI2L:
1657 return;
1658
1659 // AddNode::IdealIL can do this transform (and similar other ones):
1660 // Convert "a*b+a*c into a*(b+c)
1661 // The example had AddI(MulI(a, b), MulI(a, c)). Why did this not happen
1662 // during IGVN? There was a mutation for one of the MulI, and only
1663 // after that the pattern was as needed for the optimization. The MulI
1664 // was added to the IGVN worklist, but not the AddI. This probably
1665 // can be fixed by adding the correct pattern in add_users_of_use_to_worklist.
1666 //
1667 // Found with:
1668 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1669 // -XX:VerifyIterativeGVN=1110
1670 case Op_AddI:
1671 return;
1672
1673 // ArrayCopyNode::Ideal
1674 // calls ArrayCopyNode::prepare_array_copy
1675 // calls Compile::conv_I2X_index -> is called with sizetype = intcon(0), I think that
1676 // is not expected, and we create a range int:0..-1
1677 // calls Compile::constrained_convI2L -> creates ConvI2L(intcon(1), int:0..-1)
1678 // note: the type is already empty!
1679 // calls PhaseIterGVN::transform
1680 // calls PhaseIterGVN::transform_old
1681 // calls PhaseIterGVN::subsume_node -> subsume ConvI2L with TOP
1682 // calls Unique_Node_List::push -> pushes TOP to worklist
1683 //
1684 // Once we get back to ArrayCopyNode::prepare_array_copy, we get back TOP, and
1685 // return false. This means we eventually return nullptr from ArrayCopyNode::Ideal.
1686 //
1687 // Question: is it ok to push anything to the worklist during ::Ideal, if we will
1688 // return nullptr, indicating nothing happened?
1689 // Is it smart to do transform in Compile::constrained_convI2L, and then
1690 // check for TOP in calls ArrayCopyNode::prepare_array_copy?
1691 // Should we just allow TOP to land on the worklist, as an exception?
1692 //
1693 // Found with:
1694 // compiler/arraycopy/TestArrayCopyAsLoadsStores.java
1695 // -XX:VerifyIterativeGVN=1110
1696 case Op_ArrayCopy:
1697 return;
1698
1699 // CastLLNode::Ideal
1700 // calls ConstraintCastNode::optimize_integer_cast -> pushes CastLL through SubL
1701 //
1702 // Could be a notification issue, where updates inputs of CastLL do not notify
1703 // down through SubL to CastLL.
1704 //
1705 // Found With:
1706 // compiler/c2/TestMergeStoresMemorySegment.java#byte-array
1707 // -XX:VerifyIterativeGVN=1110
1708 case Op_CastLL:
1709 return;
1710
1711 // Similar case happens to CastII
1712 //
1713 // Found With:
1714 // compiler/c2/TestScalarReplacementMaxLiveNodes.java
1715 // -XX:VerifyIterativeGVN=1110
1716 case Op_CastII:
1717 return;
1718
1719 // MaxLNode::Ideal
1720 // calls AddNode::Ideal
1721 // calls commute -> decides to swap edges
1722 //
1723 // Another notification issue, because we check inputs of inputs?
1724 // MaxL -> Phi -> Loop
1725 // MaxL -> Phi -> MaxL
1726 //
1727 // Found with:
1728 // compiler/c2/irTests/TestIfMinMax.java
1729 // -XX:VerifyIterativeGVN=1110
1730 case Op_MaxL:
1731 case Op_MinL:
1732 return;
1733
1734 // OrINode::Ideal
1735 // calls AddNode::Ideal
1736 // calls commute -> left is Load, right not -> commute.
1737 //
1738 // Not sure why notification does not work here, seems like
1739 // the depth is only 1, so it should work. Needs investigation.
1740 //
1741 // Found with:
1742 // compiler/codegen/TestCharVect2.java#id0
1743 // -XX:VerifyIterativeGVN=1110
1744 case Op_OrI:
1745 case Op_OrL:
1746 return;
1747
1748 // Bool -> constant folded to 1.
1749 // Issue with notification?
1750 //
1751 // Found with:
1752 // compiler/c2/irTests/TestVectorizationMismatchedAccess.java
1753 // -XX:VerifyIterativeGVN=1110
1754 case Op_Bool:
1755 return;
1756
1757 // LShiftLNode::Ideal
1758 // Looks at pattern: "(x + x) << c0", converts it to "x << (c0 + 1)"
1759 // Probably a notification issue.
1760 //
1761 // Found with:
1762 // compiler/conversions/TestMoveConvI2LOrCastIIThruAddIs.java
1763 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1764 case Op_LShiftL:
1765 return;
1766
1767 // LShiftINode::Ideal
1768 // pattern: ((x + con1) << con2) -> x << con2 + con1 << con2
1769 // Could be issue with notification of inputs of inputs
1770 //
1771 // Side-note: should cases like these not be shared between
1772 // LShiftI and LShiftL?
1773 //
1774 // Found with:
1775 // compiler/escapeAnalysis/Test6689060.java
1776 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1777 case Op_LShiftI:
1778 return;
1779
1780 // AddPNode::Ideal seems to do set_req without removing lock first.
1781 // Found with various vector tests tier1-tier3.
1782 case Op_AddP:
1783 return;
1784
1785 // StrIndexOfNode::Ideal
1786 // Found in tier1-3.
1787 case Op_StrIndexOf:
1788 case Op_StrIndexOfChar:
1789 return;
1790
1791 // StrEqualsNode::Identity
1792 //
1793 // Found (linux x64 only?) with:
1794 // serviceability/sa/ClhsdbThreadContext.java
1795 // -XX:+UnlockExperimentalVMOptions -XX:LockingMode=1 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1796 // Note: The -XX:LockingMode option is not available anymore.
1797 case Op_StrEquals:
1798 return;
1799
1800 // AryEqNode::Ideal
1801 // Not investigated. Reshapes itself and adds lots of nodes to the worklist.
1802 //
1803 // Found with:
1804 // vmTestbase/vm/mlvm/meth/stress/compiler/i2c_c2i/Test.java
1805 // -XX:+UnlockDiagnosticVMOptions -XX:-TieredCompilation -XX:+StressUnstableIfTraps -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1806 case Op_AryEq:
1807 return;
1808
1809 // MergeMemNode::Ideal
1810 // Found in tier1-3. Did not investigate further yet.
1811 case Op_MergeMem:
1812 return;
1813
1814 // CMoveINode::Ideal
1815 // Found in tier1-3. Did not investigate further yet.
1816 case Op_CMoveI:
1817 return;
1818
1819 // CmpPNode::Ideal calls isa_const_java_mirror
1820 // and generates new constant nodes, even if no progress is made.
1821 // We can probably rewrite this so that only types are generated.
1822 // It seems that object types are not hashed, we could investigate
1823 // if that is an option as well.
1824 //
1825 // Found with:
1826 // java -XX:VerifyIterativeGVN=1110 -Xcomp --version
1827 case Op_CmpP:
1828 return;
1829
1830 // MinINode::Ideal
1831 // Did not investigate, but there are some patterns that might
1832 // need more notification.
1833 case Op_MinI:
1834 case Op_MaxI: // preemptively removed it as well.
1835 return;
1836 }
1837
1838 if (n->is_Store()) {
1839 // StoreNode::Ideal can do this:
1840 // // Capture an unaliased, unconditional, simple store into an initializer.
1841 // // Or, if it is independent of the allocation, hoist it above the allocation.
1842 // That replaces the Store with a MergeMem.
1843 //
1844 // We have to investigate why this does not happen during IGVN in this case.
1845 // There could also be other issues - I did not investigate further yet.
1846 //
1847 // Found with:
1848 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1849 return;
1850 }
1851
1852 if (n->is_Vector()) {
1853 // VectorNode::Ideal swaps edges, but only for ops
1854 // that are deemed commutable. But swap_edges
1855 // requires the hash to be invariant when the edges
1856 // are swapped, which is not implemented for these
1857 // vector nodes. This seems not to create any trouble
1858 // usually, but we can also get graphs where in the
1859 // end the nodes are not all commuted, so there is
1860 // definitively an issue here.
1861 //
1862 // Probably we have two options: kill the hash, or
1863 // properly make the hash commutation friendly.
1864 //
1865 // Found with:
1866 // compiler/vectorapi/TestMaskedMacroLogicVector.java
1867 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -XX:+UseParallelGC -XX:+UseNUMA
1868 return;
1869 }
1870
1871 if (n->is_Region()) {
1872 // LoopNode::Ideal calls RegionNode::Ideal.
1873 // CountedLoopNode::Ideal calls RegionNode::Ideal too.
1874 // But I got an issue because RegionNode::optimize_trichotomy
1875 // then modifies another node, and pushes nodes to the worklist
1876 // Not sure if this is ok, modifying another node like that.
1877 // Maybe it is, then we need to look into what to do with
1878 // the nodes that are now on the worklist, maybe just clear
1879 // them out again. But maybe modifying other nodes like that
1880 // is also bad design. In the end, we return nullptr for
1881 // the current CountedLoop. But the extra nodes on the worklist
1882 // trip the asserts later on.
1883 //
1884 // Found with:
1885 // compiler/eliminateAutobox/TestShortBoxing.java
1886 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1887 return;
1888 }
1889
1890 if (n->is_CallJava()) {
1891 // CallStaticJavaNode::Ideal
1892 // Led to a crash:
1893 // assert((is_CallStaticJava() && cg->is_mh_late_inline()) || (is_CallDynamicJava() && cg->is_virtual_late_inline())) failed: mismatch
1894 //
1895 // Did not investigate yet, could be a bug.
1896 // Or maybe it does not expect to be called during verification.
1897 //
1898 // Found with:
1899 // test/jdk/jdk/incubator/vector/VectorRuns.java
1900 // -XX:VerifyIterativeGVN=1110
1901
1902 // CallDynamicJavaNode::Ideal, and I think also for CallStaticJavaNode::Ideal
1903 // and possibly their subclasses.
1904 // During late inlining it can call CallJavaNode::register_for_late_inline
1905 // That means we do more rounds of late inlining, but might fail.
1906 // Then we do IGVN again, and register the node again for late inlining.
1907 // This creates an endless cycle. Everytime we try late inlining, we
1908 // are also creating more nodes, especially SafePoint and MergeMem.
1909 // These nodes are immediately rejected when the inlining fails in the
1910 // do_late_inline_check, but they still grow the memory, until we hit
1911 // the MemLimit and crash.
1912 // The assumption here seems that CallDynamicJavaNode::Ideal does not get
1913 // called repeatedly, and eventually we terminate. I fear this is not
1914 // a great assumption to make. We should investigate more.
1915 //
1916 // Found with:
1917 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-U
1918 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1919 return;
1920 }
1921
1922 // Ideal should not make progress if it returns nullptr.
1923 // We use made_progress() rather than unique() or live_nodes() because some
1924 // Ideal implementations speculatively create nodes and kill them before
1925 // returning nullptr (e.g. split_if clones a Cmp to check is_canonical).
1926 // unique() is a high-water mark that is not decremented by remove_dead_node,
1927 // so it would cause false-positives. live_nodes() accounts for dead nodes but can
1928 // decrease when Ideal removes existing nodes as side effects.
1929 // made_progress() precisely tracks meaningful transforms, and speculative
1930 // work killed via NodeOrigin::Speculative does not increment it.
1931 uint old_progress = made_progress();
1932 // The hash of a node should not change, this would indicate different inputs
1933 uint old_hash = n->hash();
1934 // Remove 'n' from hash table in case it gets modified. We want to avoid
1935 // hitting the "Need to remove from hash before changing edges" assert if
1936 // a change occurs. Instead, we would like to proceed with the optimization,
1937 // return and finally hit the assert in PhaseIterGVN::verify_optimize to get
1938 // a more meaningful message
1939 _table.hash_delete(n);
1940 Node* i = n->Ideal(this, can_reshape);
1941 // If there was no new Idealization, we are probably happy.
1942 if (i == nullptr) {
1943 uint progress = made_progress() - old_progress;
1944 if (progress != 0) {
1945 stringStream ss; // Print as a block without tty lock.
1946 ss.cr();
1947 ss.print_cr("Ideal optimization did not make progress but had side effects.");
1948 ss.print_cr(" %u transforms made progress", progress);
1949 n->dump_bfs(1, nullptr, "", &ss);
1950 tty->print_cr("%s", ss.as_string());
1951 assert(false, "Unexpected side effects from applying Ideal optimization on %s", n->Name());
1952 }
1953
1954 if (old_hash != n->hash()) {
1955 stringStream ss; // Print as a block without tty lock.
1956 ss.cr();
1957 ss.print_cr("Ideal optimization did not make progress but node hash changed.");
1958 ss.print_cr(" old_hash = %d, hash = %d", old_hash, n->hash());
1959 n->dump_bfs(1, nullptr, "", &ss);
1960 tty->print_cr("%s", ss.as_string());
1961 assert(false, "Unexpected hash change from applying Ideal optimization on %s", n->Name());
1962 }
1963
1964 // Some nodes try to push itself back to the worklist if can_reshape is
1965 // false
1966 if (!can_reshape && _worklist.size() > 0 && _worklist.pop() != n) {
1967 stringStream ss;
1968 ss.cr();
1969 ss.print_cr("Previously optimized:");
1970 n->dump_bfs(1, nullptr, "", &ss);
1971 tty->print_cr("%s", ss.as_string());
1972 assert(false, "should only push itself on worklist");
1973 }
1974 verify_empty_worklist(n);
1975
1976 // Everything is good.
1977 hash_find_insert(n);
1978 return;
1979 }
1980
1981 // We just saw a new Idealization which was not done during IGVN.
1982 stringStream ss; // Print as a block without tty lock.
1983 ss.cr();
1984 ss.print_cr("Missed Ideal optimization (can_reshape=%s):", can_reshape ? "true": "false");
1985 if (i == n) {
1986 ss.print_cr("The node was reshaped by Ideal.");
1987 } else {
1988 ss.print_cr("The node was replaced by Ideal.");
1989 ss.print_cr("Old node:");
1990 n->dump_bfs(1, nullptr, "", &ss);
1991 }
1992 ss.print_cr("The result after Ideal:");
1993 i->dump_bfs(1, nullptr, "", &ss);
1994 tty->print_cr("%s", ss.as_string());
1995
1996 assert(false, "Missed Ideal optimization opportunity in PhaseIterGVN for %s", n->Name());
1997 }
1998
1999 // Check that all Identity optimizations that could be done were done.
2000 // Asserts if it found missed optimization opportunities, and
2001 // returns normally otherwise (no missed optimization, or skipped verification).
2002 void PhaseIterGVN::verify_Identity_for(Node* n) {
2003 // First, we check a list of exceptions, where we skip verification,
2004 // because there are known cases where Ideal can optimize after IGVN.
2005 // Some may be expected and cannot be fixed, and others should be fixed.
2006 switch (n->Opcode()) {
2007 // SafePointNode::Identity can remove SafePoints, but wants to wait until
2008 // after loopopts:
2009 // // Transforming long counted loops requires a safepoint node. Do not
2010 // // eliminate a safepoint until loop opts are over.
2011 // if (in(0)->is_Proj() && !phase->C->major_progress()) {
2012 //
2013 // I think the check for major_progress does delay it until after loopopts
2014 // but it does not ensure that the node is on the IGVN worklist after
2015 // loopopts. I think we should try to instead check for
2016 // phase->C->post_loop_opts_phase() and call record_for_post_loop_opts_igvn.
2017 //
2018 // Found with:
2019 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
2020 case Op_SafePoint:
2021 return;
2022
2023 // MergeMemNode::Identity replaces the MergeMem with its base_memory if it
2024 // does not record any other memory splits.
2025 //
2026 // I did not deeply investigate, but it looks like MergeMemNode::Identity
2027 // never got called during IGVN for this node, investigate why.
2028 //
2029 // Found with:
2030 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
2031 case Op_MergeMem:
2032 return;
2033
2034 // ConstraintCastNode::Identity finds casts that are the same, except that
2035 // the control is "higher up", i.e. dominates. The call goes via
2036 // ConstraintCastNode::dominating_cast to PhaseGVN::is_dominator_helper,
2037 // which traverses up to 100 idom steps. If anything gets optimized somewhere
2038 // away from the cast, but within 100 idom steps, the cast may not be
2039 // put on the IGVN worklist any more.
2040 //
2041 // Found with:
2042 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
2043 case Op_CastPP:
2044 case Op_CastII:
2045 case Op_CastLL:
2046 return;
2047
2048 // Same issue for CheckCastPP, uses ConstraintCastNode::Identity and
2049 // checks dominator, which may be changed, but too far up for notification
2050 // to work.
2051 //
2052 // Found with:
2053 // compiler/c2/irTests/TestSkeletonPredicates.java
2054 // -XX:VerifyIterativeGVN=1110
2055 case Op_CheckCastPP:
2056 return;
2057
2058 // In SubNode::Identity, we do:
2059 // Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y
2060 // In the example, the AddI had an input replaced, the AddI is
2061 // added to the IGVN worklist, but the SubI is one link further
2062 // down and is not added. I checked add_users_of_use_to_worklist
2063 // where I would expect the SubI would be added, and I cannot
2064 // find the pattern, only this one:
2065 // If changed AddI/SubI inputs, check CmpU for range check optimization.
2066 //
2067 // Fix this "notification" issue and check if there are any other
2068 // issues.
2069 //
2070 // Found with:
2071 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
2072 case Op_SubI:
2073 case Op_SubL:
2074 return;
2075
2076 // PhiNode::Identity checks for patterns like:
2077 // r = (x != con) ? x : con;
2078 // that can be constant folded to "x".
2079 //
2080 // Call goes through PhiNode::is_cmove_id and CMoveNode::is_cmove_id.
2081 // I suspect there was some earlier change to one of the inputs, but
2082 // not all relevant outputs were put on the IGVN worklist.
2083 //
2084 // Found with:
2085 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
2086 // -XX:VerifyIterativeGVN=1110
2087 case Op_Phi:
2088 return;
2089
2090 // ConvI2LNode::Identity does
2091 // convert I2L(L2I(x)) => x
2092 //
2093 // Investigate why this did not already happen during IGVN.
2094 //
2095 // Found with:
2096 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-A
2097 // -XX:VerifyIterativeGVN=1110
2098 case Op_ConvI2L:
2099 return;
2100
2101 // AbsINode::Identity
2102 // Not investigated yet.
2103 case Op_AbsI:
2104 return;
2105 }
2106
2107 if (n->is_Load()) {
2108 // LoadNode::Identity tries to look for an earlier store value via
2109 // can_see_stored_value. I found an example where this led to
2110 // an Allocation, where we could assume the value was still zero.
2111 // So the LoadN can be replaced with a zerocon.
2112 //
2113 // Investigate why this was not already done during IGVN.
2114 // A similar issue happens with Ideal.
2115 //
2116 // Found with:
2117 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
2118 return;
2119 }
2120
2121 if (n->is_Store()) {
2122 // StoreNode::Identity
2123 // Not investigated, but found missing optimization for StoreI.
2124 // Looks like a StoreI is replaced with an InitializeNode.
2125 //
2126 // Found with:
2127 // applications/ctw/modules/java_base_2.java
2128 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -Djava.awt.headless=true -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
2129 return;
2130 }
2131
2132 if (n->is_Vector()) {
2133 // Found with tier1-3. Not investigated yet.
2134 // The observed issue was with AndVNode::Identity
2135 return;
2136 }
2137
2138 Node* i = n->Identity(this);
2139 // If we cannot find any other Identity, we are happy.
2140 if (i == n) {
2141 verify_empty_worklist(n);
2142 return;
2143 }
2144
2145 // The verification just found a new Identity that was not found during IGVN.
2146 stringStream ss; // Print as a block without tty lock.
2147 ss.cr();
2148 ss.print_cr("Missed Identity optimization:");
2149 ss.print_cr("Old node:");
2150 n->dump_bfs(1, nullptr, "", &ss);
2151 ss.print_cr("New node:");
2152 i->dump_bfs(1, nullptr, "", &ss);
2153 tty->print_cr("%s", ss.as_string());
2154
2155 assert(false, "Missed Identity optimization opportunity in PhaseIterGVN for %s", n->Name());
2156 }
2157
2158 // Some other verifications that are not specific to a particular transformation.
2159 void PhaseIterGVN::verify_node_invariants_for(const Node* n) {
2160 if (n->is_AddP()) {
2161 if (!n->as_AddP()->address_input_has_same_base()) {
2162 stringStream ss; // Print as a block without tty lock.
2163 ss.cr();
2164 ss.print_cr("Base pointers must match for AddP chain:");
2165 n->dump_bfs(2, nullptr, "", &ss);
2166 tty->print_cr("%s", ss.as_string());
2167
2168 assert(false, "Broken node invariant for %s", n->Name());
2169 }
2170 }
2171 }
2172 #endif
2173
2174 /**
2175 * Register a new node with the optimizer. Update the types array, the def-use
2176 * info. Put on worklist.
2177 */
2178 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
2179 set_type_bottom(n);
2180 _worklist.push(n);
2181 if (orig != nullptr) C->copy_node_notes_to(n, orig);
2182 return n;
2183 }
2184
2185 //------------------------------transform--------------------------------------
2186 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
2187 Node *PhaseIterGVN::transform( Node *n ) {
2188 // If brand new node, make space in type array, and give it a type.
2189 ensure_type_or_null(n);
2190 if (type_or_null(n) == nullptr) {
2191 set_type_bottom(n);
2192 }
2193
2194 if (_delay_transform) {
2195 // Add the node to the worklist but don't optimize for now
2196 _worklist.push(n);
2197 return n;
2198 }
2199
2200 return transform_old(n);
2201 }
2202
2203 Node *PhaseIterGVN::transform_old(Node* n) {
2204 NOT_PRODUCT(set_transforms());
2205 // Remove 'n' from hash table in case it gets modified
2206 _table.hash_delete(n);
2207 #ifdef ASSERT
2208 if (is_verify_def_use()) {
2209 assert(!_table.find_index(n->_idx), "found duplicate entry in table");
2210 }
2211 #endif
2212
2213 // Allow Bool -> Cmp idealisation in late inlining intrinsics that return a bool
2214 if (n->is_Cmp()) {
2215 add_users_to_worklist(n);
2216 }
2217
2218 // Apply the Ideal call in a loop until it no longer applies
2219 Node* k = n;
2220 DEBUG_ONLY(dead_loop_check(k);)
2221 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
2222 C->remove_modified_node(k);
2223 #ifndef PRODUCT
2224 uint hash_before = is_verify_Ideal_return() ? k->hash() : 0;
2225 #endif
2226 Node* i = apply_ideal(k, /*can_reshape=*/true);
2227 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
2228 #ifndef PRODUCT
2229 if (is_verify_Ideal_return()) {
2230 assert(k->outcnt() == 0 || i != nullptr || hash_before == k->hash(), "hash changed after Ideal returned nullptr for %s", k->Name());
2231 }
2232 verify_step(k);
2233 #endif
2234
2235 DEBUG_ONLY(uint loop_count = 1;)
2236 if (i != nullptr) {
2237 set_progress();
2238 }
2239 while (i != nullptr) {
2240 #ifdef ASSERT
2241 if (loop_count >= K + C->live_nodes()) {
2242 dump_infinite_loop_info(i, "PhaseIterGVN::transform_old");
2243 }
2244 #endif
2245 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
2246 // Made a change; put users of original Node on worklist
2247 add_users_to_worklist(k);
2248 // Replacing root of transform tree?
2249 if (k != i) {
2250 // Make users of old Node now use new.
2251 subsume_node(k, i);
2252 k = i;
2253 }
2254 DEBUG_ONLY(dead_loop_check(k);)
2255 // Try idealizing again
2256 DEBUG_ONLY(is_new = (k->outcnt() == 0);)
2257 C->remove_modified_node(k);
2258 #ifndef PRODUCT
2259 uint hash_before = is_verify_Ideal_return() ? k->hash() : 0;
2260 #endif
2261 i = apply_ideal(k, /*can_reshape=*/true);
2262 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
2263 #ifndef PRODUCT
2264 if (is_verify_Ideal_return()) {
2265 assert(k->outcnt() == 0 || i != nullptr || hash_before == k->hash(), "hash changed after Ideal returned nullptr for %s", k->Name());
2266 }
2267 verify_step(k);
2268 #endif
2269 DEBUG_ONLY(loop_count++;)
2270 }
2271
2272 // If brand new node, make space in type array.
2273 ensure_type_or_null(k);
2274
2275 // See what kind of values 'k' takes on at runtime
2276 const Type* t = k->Value(this);
2277 assert(t != nullptr, "value sanity");
2278
2279 // Since I just called 'Value' to compute the set of run-time values
2280 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
2281 // cache Value. Later requests for the local phase->type of this Node can
2282 // use the cached Value instead of suffering with 'bottom_type'.
2283 if (type_or_null(k) != t) {
2284 NOT_PRODUCT(inc_new_values();)
2285 set_progress();
2286 set_type(k, t);
2287 // If k is a TypeNode, capture any more-precise type permanently into Node
2288 k->raise_bottom_type(t);
2289 // Move users of node to worklist
2290 add_users_to_worklist(k);
2291 }
2292 // If 'k' computes a constant, replace it with a constant
2293 if (t->singleton() && !k->is_Con()) {
2294 set_progress();
2295 Node* con = makecon(t); // Make a constant
2296 add_users_to_worklist(k);
2297 subsume_node(k, con); // Everybody using k now uses con
2298 return con;
2299 }
2300
2301 // Now check for Identities
2302 i = k->Identity(this); // Look for a nearby replacement
2303 if (i != k) { // Found? Return replacement!
2304 set_progress();
2305 add_users_to_worklist(k);
2306 subsume_node(k, i); // Everybody using k now uses i
2307 return i;
2308 }
2309
2310 // Global Value Numbering
2311 i = hash_find_insert(k); // Check for pre-existing node
2312 if (i && (i != k)) {
2313 // Return the pre-existing node if it isn't dead
2314 set_progress();
2315 add_users_to_worklist(k);
2316 subsume_node(k, i); // Everybody using k now uses i
2317 return i;
2318 }
2319
2320 // Return Idealized original
2321 return k;
2322 }
2323
2324 //---------------------------------saturate------------------------------------
2325 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
2326 const Type* limit_type) const {
2327 return new_type->narrow(old_type);
2328 }
2329
2330 //------------------------------remove_globally_dead_node----------------------
2331 // Kill a globally dead Node. All uses are also globally dead and are
2332 // aggressively trimmed.
2333 void PhaseIterGVN::remove_globally_dead_node(Node* dead, NodeOrigin origin) {
2334 enum DeleteProgress {
2335 PROCESS_INPUTS,
2336 PROCESS_OUTPUTS
2337 };
2338 ResourceMark rm;
2339 Node_Stack stack(32);
2340 stack.push(dead, PROCESS_INPUTS);
2341
2342 while (stack.is_nonempty()) {
2343 dead = stack.node();
2344 if (dead->Opcode() == Op_SafePoint) {
2345 dead->as_SafePoint()->disconnect_from_root(this);
2346 }
2347 uint progress_state = stack.index();
2348 assert(dead != C->root(), "killing root, eh?");
2349 assert(!dead->is_top(), "add check for top when pushing");
2350 if (progress_state == PROCESS_INPUTS) {
2351 // After following inputs, continue to outputs
2352 stack.set_index(PROCESS_OUTPUTS);
2353 if (!dead->is_Con()) { // Don't kill cons but uses
2354 if (origin != NodeOrigin::Speculative) {
2355 set_progress();
2356 }
2357 bool recurse = false;
2358 // Remove from hash table
2359 _table.hash_delete( dead );
2360 // Smash all inputs to 'dead', isolating him completely
2361 for (uint i = 0; i < dead->req(); i++) {
2362 Node *in = dead->in(i);
2363 if (in != nullptr && in != C->top()) { // Points to something?
2364 int nrep = dead->replace_edge(in, nullptr, this); // Kill edges
2365 assert((nrep > 0), "sanity");
2366 if (in->outcnt() == 0) { // Made input go dead?
2367 stack.push(in, PROCESS_INPUTS); // Recursively remove
2368 recurse = true;
2369 } else if (in->outcnt() == 1 &&
2370 in->has_special_unique_user()) {
2371 _worklist.push(in->unique_out());
2372 } else if (in->outcnt() <= 2 && dead->is_Phi()) {
2373 if (in->Opcode() == Op_Region) {
2374 _worklist.push(in);
2375 } else if (in->is_Store()) {
2376 DUIterator_Fast imax, i = in->fast_outs(imax);
2377 _worklist.push(in->fast_out(i));
2378 i++;
2379 if (in->outcnt() == 2) {
2380 _worklist.push(in->fast_out(i));
2381 i++;
2382 }
2383 assert(!(i < imax), "sanity");
2384 }
2385 } else if (dead->is_data_proj_of_pure_function(in)) {
2386 _worklist.push(in);
2387 } else {
2388 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in);
2389 }
2390 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
2391 in->is_Proj() && in->in(0) != nullptr && in->in(0)->is_Initialize()) {
2392 // A Load that directly follows an InitializeNode is
2393 // going away. The Stores that follow are candidates
2394 // again to be captured by the InitializeNode.
2395 add_users_to_worklist_if(_worklist, in, [](Node* n) { return n->is_Store(); });
2396 }
2397 } // if (in != nullptr && in != C->top())
2398 } // for (uint i = 0; i < dead->req(); i++)
2399 if (recurse) {
2400 continue;
2401 }
2402 } // if (!dead->is_Con())
2403 } // if (progress_state == PROCESS_INPUTS)
2404
2405 // Aggressively kill globally dead uses
2406 // (Rather than pushing all the outs at once, we push one at a time,
2407 // plus the parent to resume later, because of the indefinite number
2408 // of edge deletions per loop trip.)
2409 if (dead->outcnt() > 0) {
2410 // Recursively remove output edges
2411 stack.push(dead->raw_out(0), PROCESS_INPUTS);
2412 } else {
2413 // Finished disconnecting all input and output edges.
2414 stack.pop();
2415 // Remove dead node from iterative worklist
2416 _worklist.remove(dead);
2417 C->remove_useless_node(dead);
2418 }
2419 } // while (stack.is_nonempty())
2420 }
2421
2422 //------------------------------subsume_node-----------------------------------
2423 // Remove users from node 'old' and add them to node 'nn'.
2424 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
2425 if (old->Opcode() == Op_SafePoint) {
2426 old->as_SafePoint()->disconnect_from_root(this);
2427 }
2428 assert( old != hash_find(old), "should already been removed" );
2429 assert( old != C->top(), "cannot subsume top node");
2430 // Copy debug or profile information to the new version:
2431 C->copy_node_notes_to(nn, old);
2432 // Move users of node 'old' to node 'nn'
2433 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
2434 Node* use = old->last_out(i); // for each use...
2435 // use might need re-hashing (but it won't if it's a new node)
2436 rehash_node_delayed(use);
2437 // Update use-def info as well
2438 // We remove all occurrences of old within use->in,
2439 // so as to avoid rehashing any node more than once.
2440 // The hash table probe swamps any outer loop overhead.
2441 uint num_edges = 0;
2442 for (uint jmax = use->len(), j = 0; j < jmax; j++) {
2443 if (use->in(j) == old) {
2444 use->set_req(j, nn);
2445 ++num_edges;
2446 }
2447 }
2448 i -= num_edges; // we deleted 1 or more copies of this edge
2449 }
2450
2451 // Search for instance field data PhiNodes in the same region pointing to the old
2452 // memory PhiNode and update their instance memory ids to point to the new node.
2453 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != nullptr) {
2454 Node* region = old->in(0);
2455 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2456 PhiNode* phi = region->fast_out(i)->isa_Phi();
2457 if (phi != nullptr && phi->inst_mem_id() == (int)old->_idx) {
2458 phi->set_inst_mem_id((int)nn->_idx);
2459 }
2460 }
2461 }
2462
2463 // Smash all inputs to 'old', isolating him completely
2464 Node *temp = new Node(1);
2465 temp->init_req(0,nn); // Add a use to nn to prevent him from dying
2466 remove_dead_node(old, NodeOrigin::Graph);
2467 temp->del_req(0); // Yank bogus edge
2468 if (nn != nullptr && nn->outcnt() == 0) {
2469 _worklist.push(nn);
2470 }
2471 #ifndef PRODUCT
2472 if (is_verify_def_use()) {
2473 for ( int i = 0; i < _verify_window_size; i++ ) {
2474 if ( _verify_window[i] == old )
2475 _verify_window[i] = nn;
2476 }
2477 }
2478 #endif
2479 temp->destruct(this); // reuse the _idx of this little guy
2480 }
2481
2482 void PhaseIterGVN::replace_in_uses(Node* n, Node* m) {
2483 assert(n != nullptr, "sanity");
2484 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2485 Node* u = n->fast_out(i);
2486 if (u != n) {
2487 rehash_node_delayed(u);
2488 int nb = u->replace_edge(n, m);
2489 --i, imax -= nb;
2490 }
2491 }
2492 assert(n->outcnt() == 0, "all uses must be deleted");
2493 }
2494
2495 //------------------------------add_users_to_worklist--------------------------
2496 void PhaseIterGVN::add_users_to_worklist0(Node* n, Unique_Node_List& worklist) {
2497 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2498 worklist.push(n->fast_out(i)); // Push on worklist
2499 }
2500 }
2501
2502 // Return counted loop Phi if as a counted loop exit condition, cmp
2503 // compares the induction variable with n
2504 static PhiNode* countedloop_phi_from_cmp(CmpNode* cmp, Node* n) {
2505 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
2506 Node* bol = cmp->fast_out(i);
2507 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
2508 Node* iff = bol->fast_out(i2);
2509 if (iff->is_BaseCountedLoopEnd()) {
2510 BaseCountedLoopEndNode* cle = iff->as_BaseCountedLoopEnd();
2511 if (cle->limit() == n) {
2512 PhiNode* phi = cle->phi();
2513 if (phi != nullptr) {
2514 return phi;
2515 }
2516 }
2517 }
2518 }
2519 }
2520 return nullptr;
2521 }
2522
2523 void PhaseIterGVN::add_users_to_worklist(Node *n) {
2524 add_users_to_worklist0(n, _worklist);
2525
2526 Unique_Node_List& worklist = _worklist;
2527 // Move users of node to worklist
2528 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2529 Node* use = n->fast_out(i); // Get use
2530 add_users_of_use_to_worklist(n, use, worklist);
2531 }
2532 }
2533
2534 void PhaseIterGVN::add_users_of_use_to_worklist(Node* n, Node* use, Unique_Node_List& worklist) {
2535 if(use->is_Multi() || // Multi-definer? Push projs on worklist
2536 use->is_Store() ) // Enable store/load same address
2537 add_users_to_worklist0(use, worklist);
2538
2539 // If we changed the receiver type to a call, we need to revisit
2540 // the Catch following the call. It's looking for a non-null
2541 // receiver to know when to enable the regular fall-through path
2542 // in addition to the NullPtrException path.
2543 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
2544 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
2545 if (p != nullptr) {
2546 add_users_to_worklist0(p, worklist);
2547 }
2548 }
2549
2550 // AndLNode::Ideal folds GraphKit::mark_word_test patterns. Give it a chance to run.
2551 if (n->is_Load() && use->is_Phi()) {
2552 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2553 Node* u = use->fast_out(i);
2554 if (u->Opcode() == Op_AndL) {
2555 worklist.push(u);
2556 }
2557 }
2558 }
2559
2560 uint use_op = use->Opcode();
2561 if(use->is_Cmp()) { // Enable CMP/BOOL optimization
2562 add_users_to_worklist0(use, worklist); // Put Bool on worklist
2563 if (use->outcnt() > 0) {
2564 Node* bol = use->raw_out(0);
2565 if (bol->outcnt() > 0) {
2566 Node* iff = bol->raw_out(0);
2567 if (iff->outcnt() == 2) {
2568 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
2569 // phi merging either 0 or 1 onto the worklist
2570 Node* ifproj0 = iff->raw_out(0);
2571 Node* ifproj1 = iff->raw_out(1);
2572 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
2573 Node* region0 = ifproj0->raw_out(0);
2574 Node* region1 = ifproj1->raw_out(0);
2575 if( region0 == region1 )
2576 add_users_to_worklist0(region0, worklist);
2577 }
2578 }
2579 }
2580 }
2581 if (use_op == Op_CmpI || use_op == Op_CmpL) {
2582 Node* phi = countedloop_phi_from_cmp(use->as_Cmp(), n);
2583 if (phi != nullptr) {
2584 // Input to the cmp of a loop exit check has changed, thus
2585 // the loop limit may have changed, which can then change the
2586 // range values of the trip-count Phi.
2587 worklist.push(phi);
2588 }
2589 }
2590 if (use_op == Op_CmpI) {
2591 Node* cmp = use;
2592 Node* in1 = cmp->in(1);
2593 Node* in2 = cmp->in(2);
2594 // Notify CmpI / If pattern from CastIINode::Value (left pattern).
2595 // Must also notify if in1 is modified and possibly turns into X (right pattern).
2596 //
2597 // in1 in2 in1 in2
2598 // | | | |
2599 // +--- | --+ | |
2600 // | | | | |
2601 // CmpINode | CmpINode
2602 // | | |
2603 // BoolNode | BoolNode
2604 // | | OR |
2605 // IfNode | IfNode
2606 // | | |
2607 // IfProj | IfProj X
2608 // | | | |
2609 // CastIINode CastIINode
2610 //
2611 if (in1 != in2) { // if they are equal, the CmpI can fold them away
2612 if (in1 == n) {
2613 // in1 modified -> could turn into X -> do traversal based on right pattern.
2614 for (DUIterator_Fast i2max, i2 = cmp->fast_outs(i2max); i2 < i2max; i2++) {
2615 Node* bol = cmp->fast_out(i2); // For each Bool
2616 if (bol->is_Bool()) {
2617 for (DUIterator_Fast i3max, i3 = bol->fast_outs(i3max); i3 < i3max; i3++) {
2618 Node* iff = bol->fast_out(i3); // For each If
2619 if (iff->is_If()) {
2620 for (DUIterator_Fast i4max, i4 = iff->fast_outs(i4max); i4 < i4max; i4++) {
2621 Node* if_proj = iff->fast_out(i4); // For each IfProj
2622 assert(if_proj->is_IfProj(), "If only has IfTrue and IfFalse as outputs");
2623 for (DUIterator_Fast i5max, i5 = if_proj->fast_outs(i5max); i5 < i5max; i5++) {
2624 Node* castii = if_proj->fast_out(i5); // For each CastII
2625 if (castii->is_CastII() &&
2626 castii->as_CastII()->carry_dependency()) {
2627 worklist.push(castii);
2628 }
2629 }
2630 }
2631 }
2632 }
2633 }
2634 }
2635 } else {
2636 // Only in2 modified -> can assume X == in2 (left pattern).
2637 assert(n == in2, "only in2 modified");
2638 // Find all CastII with input in1.
2639 for (DUIterator_Fast jmax, j = in1->fast_outs(jmax); j < jmax; j++) {
2640 Node* castii = in1->fast_out(j);
2641 if (castii->is_CastII() && castii->as_CastII()->carry_dependency()) {
2642 // Find If.
2643 if (castii->in(0) != nullptr && castii->in(0)->in(0) != nullptr && castii->in(0)->in(0)->is_If()) {
2644 Node* ifnode = castii->in(0)->in(0);
2645 // Check that if connects to the cmp
2646 if (ifnode->in(1) != nullptr && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == cmp) {
2647 worklist.push(castii);
2648 }
2649 }
2650 }
2651 }
2652 }
2653 }
2654 }
2655 }
2656
2657 // Inline type nodes can have other inline types as users. If an input gets
2658 // updated, make sure that inline type users get a chance for optimization.
2659 if (use->is_InlineType() || use->is_DecodeN()) {
2660 auto push_the_uses_to_worklist = [&](Node* n){
2661 if (n->is_InlineType()) {
2662 worklist.push(n);
2663 }
2664 };
2665 auto is_boundary = [](Node* n){ return !n->is_InlineType(); };
2666 use->visit_uses(push_the_uses_to_worklist, is_boundary, true);
2667 }
2668 // If changed Cast input, notify down for Phi, Sub, and Xor - all do "uncast"
2669 // Patterns:
2670 // ConstraintCast+ -> Sub
2671 // ConstraintCast+ -> Phi
2672 // ConstraintCast+ -> Xor
2673 if (use->is_ConstraintCast()) {
2674 auto push_the_uses_to_worklist = [&](Node* n){
2675 if (n->is_Phi() || n->is_Sub() || n->Opcode() == Op_XorI || n->Opcode() == Op_XorL) {
2676 worklist.push(n);
2677 }
2678 };
2679 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast(); };
2680 use->visit_uses(push_the_uses_to_worklist, is_boundary);
2681 }
2682 // If changed LShift inputs, check RShift/URShift users for
2683 // "(X << C) >> C" sign-ext and "(X << C) >>> C" zero-ext optimizations.
2684 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2685 add_users_to_worklist_if(worklist, use, [](Node* u) {
2686 return u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL ||
2687 u->Opcode() == Op_URShiftI || u->Opcode() == Op_URShiftL;
2688 });
2689 }
2690 // If changed LShift inputs, check And users for shift and mask (And) operation
2691 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2692 add_users_to_worklist_if(worklist, use, [](Node* u) {
2693 return u->Opcode() == Op_AndI || u->Opcode() == Op_AndL;
2694 });
2695 }
2696 // If changed AddI/SubI inputs, check CmpU for range check optimization.
2697 if (use_op == Op_AddI || use_op == Op_SubI) {
2698 add_users_to_worklist_if(worklist, use, [](Node* u) {
2699 return u->Opcode() == Op_CmpU;
2700 });
2701 }
2702 // If changed AddI/AddL inputs, check URShift users for
2703 // "((X << z) + Y) >>> z" optimization in URShift{I,L}Node::Ideal.
2704 if (use_op == Op_AddI || use_op == Op_AddL) {
2705 add_users_to_worklist_if(worklist, use, [](Node* u) {
2706 return u->Opcode() == Op_URShiftI || u->Opcode() == Op_URShiftL;
2707 });
2708 }
2709 // If changed LShiftI/LShiftL inputs, check AddI/AddL users for their
2710 // URShiftI/URShiftL users for "((x << z) + y) >>> z" optimization opportunity
2711 // (see URShiftINode::Ideal). Handles the case where the LShift input changes.
2712 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2713 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2714 Node* add = use->fast_out(i2);
2715 if (add->Opcode() == Op_AddI || add->Opcode() == Op_AddL) {
2716 add_users_to_worklist_if(worklist, add, [](Node* u) {
2717 return u->Opcode() == Op_URShiftI || u->Opcode() == Op_URShiftL;
2718 });
2719 }
2720 }
2721 }
2722 // If changed AndI/AndL inputs, check RShift/URShift users for "(x & mask) >> shift" optimization opportunity
2723 if (use_op == Op_AndI || use_op == Op_AndL) {
2724 add_users_to_worklist_if(worklist, use, [](Node* u) {
2725 return u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL ||
2726 u->Opcode() == Op_URShiftI || u->Opcode() == Op_URShiftL;
2727 });
2728 }
2729 // Check for redundant conversion patterns:
2730 // ConvD2L->ConvL2D->ConvD2L
2731 // ConvF2I->ConvI2F->ConvF2I
2732 // ConvF2L->ConvL2F->ConvF2L
2733 // ConvI2F->ConvF2I->ConvI2F
2734 // Note: there may be other 3-nodes conversion chains that would require to be added here, but these
2735 // are the only ones that are known to trigger missed optimizations otherwise
2736 if (use_op == Op_ConvL2D ||
2737 use_op == Op_ConvI2F ||
2738 use_op == Op_ConvL2F ||
2739 use_op == Op_ConvF2I) {
2740 add_users_to_worklist_if(worklist, use, [=](Node* u) {
2741 return (use_op == Op_ConvL2D && u->Opcode() == Op_ConvD2L) ||
2742 (use_op == Op_ConvI2F && u->Opcode() == Op_ConvF2I) ||
2743 (use_op == Op_ConvL2F && u->Opcode() == Op_ConvF2L) ||
2744 (use_op == Op_ConvF2I && u->Opcode() == Op_ConvI2F);
2745 });
2746 }
2747 // ConvD2F::Ideal matches ConvD2F(SqrtD(ConvF2D(x))) => SqrtF(x).
2748 // Notify ConvD2F users of SqrtD when any input of the SqrtD changes.
2749 if (use_op == Op_SqrtD) {
2750 add_users_to_worklist_if(worklist, use, [](Node* u) { return u->Opcode() == Op_ConvD2F; });
2751 }
2752 // ConvF2HF::Ideal matches ConvF2HF(binopF(ConvHF2F(...))) => FP16BinOp(...).
2753 // Notify ConvF2HF users of float binary ops when any input changes.
2754 if (Float16NodeFactory::is_float32_binary_oper(use_op)) {
2755 add_users_to_worklist_if(worklist, use, [](Node* u) { return u->Opcode() == Op_ConvF2HF; });
2756 }
2757 // If changed AddP inputs:
2758 // - check Stores for loop invariant, and
2759 // - if the changed input is the offset, check constant-offset AddP users for
2760 // address expression flattening.
2761 if (use_op == Op_AddP) {
2762 bool offset_changed = n == use->in(AddPNode::Offset);
2763 add_users_to_worklist_if(worklist, use, [=](Node* u) {
2764 return u->is_Mem() ||
2765 (offset_changed && u->is_AddP() && u->in(AddPNode::Offset)->is_Con());
2766 });
2767 }
2768 // Check for "abs(0-x)" into "abs(x)" conversion
2769 if (use->is_Sub()) {
2770 add_users_to_worklist_if(worklist, use, [](Node* u) {
2771 return u->Opcode() == Op_AbsD || u->Opcode() == Op_AbsF ||
2772 u->Opcode() == Op_AbsL || u->Opcode() == Op_AbsI;
2773 });
2774 }
2775 // Check for Max/Min(A, Max/Min(B, C)) where A == B or A == C
2776 if (use->is_MinMax()) {
2777 add_users_to_worklist_if(worklist, use, [](Node* u) { return u->is_MinMax(); });
2778 }
2779 auto enqueue_init_mem_projs = [&](ProjNode* proj) {
2780 add_users_to_worklist0(proj, worklist);
2781 };
2782 // If changed initialization activity, check dependent Stores
2783 if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
2784 InitializeNode* init = use->as_Allocate()->initialization();
2785 if (init != nullptr) {
2786 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2787 }
2788 }
2789 // If the ValidLengthTest input changes then the fallthrough path out of the AllocateArray may have become dead.
2790 // CatchNode::Value() is responsible for killing that path. The CatchNode has to be explicitly enqueued for igvn
2791 // to guarantee the change is not missed.
2792 if (use_op == Op_AllocateArray && n == use->in(AllocateNode::ValidLengthTest)) {
2793 Node* p = use->as_AllocateArray()->proj_out_or_null(TypeFunc::Control);
2794 if (p != nullptr) {
2795 add_users_to_worklist0(p, worklist);
2796 }
2797 }
2798
2799 if (use_op == Op_Initialize) {
2800 InitializeNode* init = use->as_Initialize();
2801 init->for_each_proj(enqueue_init_mem_projs, TypeFunc::Memory);
2802 }
2803 // Loading the java mirror from a Klass requires two loads and the type
2804 // of the mirror load depends on the type of 'n'. See LoadNode::Value().
2805 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror))))
2806 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2807 bool has_load_barrier_nodes = bs->has_load_barrier_nodes();
2808
2809 if (use_op == Op_CastP2X) {
2810 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2811 Node* u = use->fast_out(i2);
2812 // TODO 8350865 Still needed? Yes, I think this is from PhaseMacroExpand::expand_mh_intrinsic_return
2813 if (u->Opcode() == Op_AndX) {
2814 worklist.push(u);
2815 }
2816 // Search for CmpL(OrL(CastP2X(..), CastP2X(..)), 0L)
2817 if (u->Opcode() == Op_OrL) {
2818 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
2819 Node* cmp = u->fast_out(i3);
2820 if (cmp->Opcode() == Op_CmpL) {
2821 worklist.push(cmp);
2822 }
2823 }
2824 }
2825 }
2826 }
2827 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
2828 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2829 Node* u = use->fast_out(i2);
2830 const Type* ut = u->bottom_type();
2831 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
2832 if (has_load_barrier_nodes) {
2833 // Search for load barriers behind the load
2834 add_users_to_worklist_if(worklist, u, [&](Node* b) {
2835 return bs->is_gc_barrier_node(b);
2836 });
2837 }
2838 worklist.push(u);
2839 }
2840 }
2841 }
2842 // Give CallStaticJavaNode::remove_useless_allocation a chance to run
2843 if (use->is_Region()) {
2844 Node* c = use;
2845 do {
2846 c = c->unique_ctrl_out_or_null();
2847 } while (c != nullptr && c->is_Region());
2848 if (c != nullptr && c->is_CallStaticJava() && c->as_CallStaticJava()->uncommon_trap_request() != 0) {
2849 worklist.push(c);
2850 }
2851 }
2852 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
2853 assert(use->outcnt() <= 1, "OpaqueZeroTripGuard can't be shared");
2854 if (use->outcnt() == 1) {
2855 Node* cmp = use->unique_out();
2856 worklist.push(cmp);
2857 }
2858 }
2859 // VectorMaskToLongNode::Ideal_MaskAll looks through VectorStoreMask
2860 // to fold constant masks.
2861 if (use_op == Op_VectorStoreMask) {
2862 add_users_to_worklist_if(worklist, use, [](Node* u) { return u->Opcode() == Op_VectorMaskToLong; });
2863 }
2864
2865 // From CastX2PNode::Ideal
2866 // CastX2P(AddX(x, y))
2867 // CastX2P(SubX(x, y))
2868 if (use->Opcode() == Op_AddX || use->Opcode() == Op_SubX) {
2869 add_users_to_worklist_if(worklist, use, [](Node* u) { return u->Opcode() == Op_CastX2P; });
2870 }
2871
2872 /* AndNode has a special handling when one of the operands is a LShiftNode:
2873 * (LHS << s) & RHS
2874 * if RHS fits in less than s bits, the value of this expression is 0.
2875 * The difficulty is that there might be a conversion node (ConvI2L) between
2876 * the LShiftINode and the AndLNode, like so:
2877 * AndLNode(ConvI2L(LShiftI(LHS, s)), RHS)
2878 * This case is handled by And[IL]Node::Value(PhaseGVN*)
2879 * (see `AndIL_min_trailing_zeros`).
2880 *
2881 * But, when the shift is updated during IGVN, pushing the user (ConvI2L)
2882 * is not enough: there might be no update happening there. We need to
2883 * directly push the And[IL]Node on the worklist, jumping over ConvI2L.
2884 *
2885 * Moreover we can have ConstraintCasts in between. It may look like
2886 * ConstraintCast+ -> ConvI2L -> ConstraintCast+ -> And
2887 * and And[IL]Node::Value(PhaseGVN*) still handles that by looking through casts.
2888 * So we must deal with that as well.
2889 */
2890 if (use->is_ConstraintCast() || use_op == Op_ConvI2L) {
2891 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast() && n->Opcode() != Op_ConvI2L; };
2892 auto push_and_to_worklist = [&worklist](Node* n){
2893 if (n->Opcode() == Op_AndL || n->Opcode() == Op_AndI) {
2894 worklist.push(n);
2895 }
2896 };
2897 use->visit_uses(push_and_to_worklist, is_boundary);
2898 }
2899
2900 // If changed Sub inputs, check Add for identity.
2901 // e.g., (x - y) + y -> x; x + (y - x) -> y.
2902 if (use_op == Op_SubI || use_op == Op_SubL) {
2903 const int add_op = (use_op == Op_SubI) ? Op_AddI : Op_AddL;
2904 add_users_to_worklist_if(worklist, use, [=](Node* u) { return u->Opcode() == add_op; });
2905 }
2906 }
2907
2908 /**
2909 * Remove the speculative part of all types that we know of
2910 */
2911 void PhaseIterGVN::remove_speculative_types() {
2912 assert(UseTypeSpeculation, "speculation is off");
2913 for (uint i = 0; i < _types.Size(); i++) {
2914 const Type* t = _types.fast_lookup(i);
2915 if (t != nullptr) {
2916 _types.map(i, t->remove_speculative());
2917 }
2918 }
2919 _table.check_no_speculative_types();
2920 }
2921
2922 //=============================================================================
2923 #ifndef PRODUCT
2924 uint PhaseCCP::_total_invokes = 0;
2925 uint PhaseCCP::_total_constants = 0;
2926 #endif
2927 //------------------------------PhaseCCP---------------------------------------
2928 // Conditional Constant Propagation, ala Wegman & Zadeck
2929 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
2930 NOT_PRODUCT( clear_constants(); )
2931 assert( _worklist.size() == 0, "" );
2932 _phase = PhaseValuesType::ccp;
2933 analyze();
2934 }
2935
2936 #ifndef PRODUCT
2937 //------------------------------~PhaseCCP--------------------------------------
2938 PhaseCCP::~PhaseCCP() {
2939 inc_invokes();
2940 _total_constants += count_constants();
2941 }
2942 #endif
2943
2944
2945 #ifdef ASSERT
2946 void PhaseCCP::verify_type(Node* n, const Type* tnew, const Type* told) {
2947 if (tnew->meet(told) != tnew->remove_speculative()) {
2948 n->dump(3);
2949 tty->print("told = "); told->dump(); tty->cr();
2950 tty->print("tnew = "); tnew->dump(); tty->cr();
2951 fatal("Not monotonic");
2952 }
2953 assert(!told->isa_int() || !tnew->isa_int() || told->is_int()->_widen <= tnew->is_int()->_widen, "widen increases");
2954 assert(!told->isa_long() || !tnew->isa_long() || told->is_long()->_widen <= tnew->is_long()->_widen, "widen increases");
2955 }
2956 #endif //ASSERT
2957
2958 // In this analysis, all types are initially set to TOP. We iteratively call Value() on all nodes of the graph until
2959 // we reach a fixed-point (i.e. no types change anymore). We start with a list that only contains the root node. Each time
2960 // a new type is set, we push all uses of that node back to the worklist (in some cases, we also push grandchildren
2961 // or nodes even further down back to the worklist because their type could change as a result of the current type
2962 // change).
2963 void PhaseCCP::analyze() {
2964 // Initialize all types to TOP, optimistic analysis
2965 for (uint i = 0; i < C->unique(); i++) {
2966 _types.map(i, Type::TOP);
2967 }
2968
2969 // CCP worklist is placed on a local arena, so that we can allow ResourceMarks on "Compile::current()->resource_arena()".
2970 // We also do not want to put the worklist on "Compile::current()->comp_arena()", as that one only gets de-allocated after
2971 // Compile is over. The local arena gets de-allocated at the end of its scope.
2972 ResourceArea local_arena(mtCompiler);
2973 Unique_Node_List worklist(&local_arena);
2974 Unique_Node_List worklist_revisit(&local_arena);
2975 DEBUG_ONLY(Unique_Node_List worklist_verify(&local_arena);)
2976
2977 // Push root onto worklist
2978 worklist.push(C->root());
2979
2980 assert(_root_and_safepoints.size() == 0, "must be empty (unused)");
2981 _root_and_safepoints.push(C->root());
2982
2983 // This is the meat of CCP: pull from worklist; compute new value; push changes out.
2984
2985 // Do the first round. Since all initial types are TOP, this will visit all alive nodes.
2986 while (worklist.size() != 0) {
2987 Node* n = fetch_next_node(worklist);
2988 DEBUG_ONLY(worklist_verify.push(n);)
2989 if (needs_revisit(n)) {
2990 worklist_revisit.push(n);
2991 }
2992 if (n->is_SafePoint()) {
2993 // Make sure safepoints are processed by PhaseCCP::transform even if they are
2994 // not reachable from the bottom. Otherwise, infinite loops would be removed.
2995 _root_and_safepoints.push(n);
2996 }
2997 analyze_step(worklist, n);
2998 }
2999
3000 // More rounds to catch updates far in the graph.
3001 // Revisit nodes that might be able to refine their types at the end of the round.
3002 // If so, process these nodes. If there is remaining work, start another round.
3003 do {
3004 while (worklist.size() != 0) {
3005 Node* n = fetch_next_node(worklist);
3006 analyze_step(worklist, n);
3007 }
3008 for (uint t = 0; t < worklist_revisit.size(); t++) {
3009 Node* n = worklist_revisit.at(t);
3010 analyze_step(worklist, n);
3011 }
3012 } while (worklist.size() != 0);
3013
3014 DEBUG_ONLY(verify_analyze(worklist_verify);)
3015 }
3016
3017 void PhaseCCP::analyze_step(Unique_Node_List& worklist, Node* n) {
3018 const Type* new_type = n->Value(this);
3019 if (new_type != type(n)) {
3020 DEBUG_ONLY(verify_type(n, new_type, type(n));)
3021 dump_type_and_node(n, new_type);
3022 set_type(n, new_type);
3023 push_child_nodes_to_worklist(worklist, n);
3024 }
3025 if (KillPathsReachableByDeadTypeNode && n->is_Type() && new_type == Type::TOP) {
3026 // Keep track of Type nodes to kill CFG paths that use Type
3027 // nodes that become dead.
3028 _maybe_top_type_nodes.push(n);
3029 }
3030 }
3031
3032 // Some nodes can refine their types due to type change somewhere deep
3033 // in the graph. We will need to revisit them before claiming convergence.
3034 // Add nodes here if particular *Node::Value is doing deep graph traversals
3035 // not handled by PhaseCCP::push_more_uses().
3036 bool PhaseCCP::needs_revisit(Node* n) const {
3037 // LoadNode performs deep traversals. Load is not notified for changes far away.
3038 if (n->is_Load()) {
3039 return true;
3040 }
3041 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
3042 if (n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
3043 return true;
3044 }
3045 return false;
3046 }
3047
3048 #ifdef ASSERT
3049 // For every node n on verify list, check if type(n) == n->Value()
3050 // Note for CCP the non-convergence can lead to unsound analysis and mis-compilation.
3051 // Therefore, we are verifying Value convergence strictly.
3052 void PhaseCCP::verify_analyze(Unique_Node_List& worklist_verify) {
3053 while (worklist_verify.size()) {
3054 Node* n = worklist_verify.pop();
3055
3056 // An assert in verify_Value_for means that PhaseCCP is not at fixpoint
3057 // and that the analysis result may be unsound.
3058 // If this happens, check why the reported nodes were not processed again in CCP.
3059 // We should either make sure that these nodes are properly added back to the CCP worklist
3060 // in PhaseCCP::push_child_nodes_to_worklist() to update their type in the same round,
3061 // or that they are added in PhaseCCP::needs_revisit() so that analysis revisits
3062 // them at the end of the round.
3063 verify_Value_for(n, true);
3064 }
3065 }
3066 #endif
3067
3068 // Fetch next node from worklist to be examined in this iteration.
3069 Node* PhaseCCP::fetch_next_node(Unique_Node_List& worklist) {
3070 if (StressCCP) {
3071 return worklist.remove(C->random() % worklist.size());
3072 } else {
3073 return worklist.pop();
3074 }
3075 }
3076
3077 #ifndef PRODUCT
3078 void PhaseCCP::dump_type_and_node(const Node* n, const Type* t) {
3079 if (TracePhaseCCP) {
3080 t->dump();
3081 do {
3082 tty->print("\t");
3083 } while (tty->position() < 16);
3084 n->dump();
3085 }
3086 }
3087 #endif
3088
3089 bool PhaseCCP::not_bottom_type(Node* n) const {
3090 return n->bottom_type() != type(n);
3091 }
3092
3093 // We need to propagate the type change of 'n' to all its uses. Depending on the kind of node, additional nodes
3094 // (grandchildren or even further down) need to be revisited as their types could also be improved as a result
3095 // of the new type of 'n'. Push these nodes to the worklist.
3096 void PhaseCCP::push_child_nodes_to_worklist(Unique_Node_List& worklist, Node* n) const {
3097 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3098 Node* use = n->fast_out(i);
3099 push_if_not_bottom_type(worklist, use);
3100 push_more_uses(worklist, n, use);
3101 }
3102 }
3103
3104 void PhaseCCP::push_if_not_bottom_type(Unique_Node_List& worklist, Node* n) const {
3105 if (not_bottom_type(n)) {
3106 worklist.push(n);
3107 }
3108 }
3109
3110 // For some nodes, we need to propagate the type change to grandchildren or even further down.
3111 // Add them back to the worklist.
3112 void PhaseCCP::push_more_uses(Unique_Node_List& worklist, Node* parent, const Node* use) const {
3113 push_phis(worklist, use);
3114 push_catch(worklist, use);
3115 push_cmpu(worklist, use);
3116 push_counted_loop_phi(worklist, parent, use);
3117 push_cast(worklist, use);
3118 push_loadp(worklist, use);
3119 push_and(worklist, parent, use);
3120 push_cast_ii(worklist, parent, use);
3121 push_opaque_zero_trip_guard(worklist, use);
3122 push_bool_with_cmpu_and_mask(worklist, use);
3123 }
3124
3125
3126 // We must recheck Phis too if use is a Region.
3127 void PhaseCCP::push_phis(Unique_Node_List& worklist, const Node* use) const {
3128 if (use->is_Region()) {
3129 add_users_to_worklist_if(worklist, use, [&](Node* u) {
3130 return not_bottom_type(u);
3131 });
3132 }
3133 }
3134
3135 // If we changed the receiver type to a call, we need to revisit the Catch node following the call. It's looking for a
3136 // non-null receiver to know when to enable the regular fall-through path in addition to the NullPtrException path.
3137 // Same is true if the type of a ValidLengthTest input to an AllocateArrayNode changes.
3138 void PhaseCCP::push_catch(Unique_Node_List& worklist, const Node* use) {
3139 if (use->is_Call()) {
3140 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3141 Node* proj = use->fast_out(i);
3142 if (proj->is_Proj() && proj->as_Proj()->_con == TypeFunc::Control) {
3143 Node* catch_node = proj->find_out_with(Op_Catch);
3144 if (catch_node != nullptr) {
3145 worklist.push(catch_node);
3146 }
3147 }
3148 }
3149 }
3150 }
3151
3152 // CmpU nodes can get their type information from two nodes up in the graph (instead of from the nodes immediately
3153 // above). Make sure they are added to the worklist if nodes they depend on are updated since they could be missed
3154 // and get wrong types otherwise.
3155 void PhaseCCP::push_cmpu(Unique_Node_List& worklist, const Node* use) const {
3156 uint use_op = use->Opcode();
3157 if (use_op == Op_AddI || use_op == Op_SubI) {
3158 // Got a CmpU or CmpU3 which might need the new type information from node n.
3159 add_users_to_worklist_if(worklist, use, [&](Node* u) {
3160 uint op = u->Opcode();
3161 return (op == Op_CmpU || op == Op_CmpU3) && not_bottom_type(u);
3162 });
3163 }
3164 }
3165
3166 // Look for the following shape, which can be optimized by BoolNode::Value_cmpu_and_mask() (i.e. corresponds to case
3167 // (1b): "(m & x) <u (m + 1))".
3168 // If any of the inputs on the level (%%) change, we need to revisit Bool because we could have prematurely found that
3169 // the Bool is constant (i.e. case (1b) can be applied) which could become invalid with new type information during CCP.
3170 //
3171 // m x m 1 (%%)
3172 // \ / \ /
3173 // AndI AddI
3174 // \ /
3175 // CmpU
3176 // |
3177 // Bool
3178 //
3179 void PhaseCCP::push_bool_with_cmpu_and_mask(Unique_Node_List& worklist, const Node* use) const {
3180 uint use_op = use->Opcode();
3181 if (use_op != Op_AndI && (use_op != Op_AddI || use->in(2)->find_int_con(0) != 1)) {
3182 // Not "m & x" or "m + 1"
3183 return;
3184 }
3185 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3186 Node* cmpu = use->fast_out(i);
3187 if (cmpu->Opcode() == Op_CmpU) {
3188 push_bool_matching_case1b(worklist, cmpu);
3189 }
3190 }
3191 }
3192
3193 // Push any Bool below 'cmpu' that matches case (1b) of BoolNode::Value_cmpu_and_mask().
3194 void PhaseCCP::push_bool_matching_case1b(Unique_Node_List& worklist, const Node* cmpu) const {
3195 assert(cmpu->Opcode() == Op_CmpU, "must be");
3196 for (DUIterator_Fast imax, i = cmpu->fast_outs(imax); i < imax; i++) {
3197 Node* bol = cmpu->fast_out(i);
3198 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::lt) {
3199 // Not a Bool with "<u"
3200 continue;
3201 }
3202 Node* andI = cmpu->in(1);
3203 Node* addI = cmpu->in(2);
3204 if (andI->Opcode() != Op_AndI || addI->Opcode() != Op_AddI || addI->in(2)->find_int_con(0) != 1) {
3205 // Not "m & x" and "m + 1"
3206 continue;
3207 }
3208
3209 Node* m = addI->in(1);
3210 if (m == andI->in(1) || m == andI->in(2)) {
3211 // Is "m" shared? Matched (1b) and thus we revisit Bool.
3212 push_if_not_bottom_type(worklist, bol);
3213 }
3214 }
3215 }
3216
3217 // If n is used in a counted loop exit condition, then the type of the counted loop's Phi depends on the type of 'n'.
3218 // Seem PhiNode::Value().
3219 void PhaseCCP::push_counted_loop_phi(Unique_Node_List& worklist, Node* parent, const Node* use) {
3220 uint use_op = use->Opcode();
3221 if (use_op == Op_CmpI || use_op == Op_CmpL) {
3222 PhiNode* phi = countedloop_phi_from_cmp(use->as_Cmp(), parent);
3223 if (phi != nullptr) {
3224 worklist.push(phi);
3225 }
3226 }
3227 }
3228
3229 // TODO 8350865 Still needed? Yes, I think this is from PhaseMacroExpand::expand_mh_intrinsic_return
3230 void PhaseCCP::push_cast(Unique_Node_List& worklist, const Node* use) {
3231 uint use_op = use->Opcode();
3232 if (use_op == Op_CastP2X) {
3233 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
3234 Node* u = use->fast_out(i2);
3235 if (u->Opcode() == Op_AndX) {
3236 worklist.push(u);
3237 }
3238 }
3239 }
3240 }
3241
3242 // Loading the java mirror from a Klass requires two loads and the type of the mirror load depends on the type of 'n'.
3243 // See LoadNode::Value().
3244 void PhaseCCP::push_loadp(Unique_Node_List& worklist, const Node* use) const {
3245 BarrierSetC2* barrier_set = BarrierSet::barrier_set()->barrier_set_c2();
3246 bool has_load_barrier_nodes = barrier_set->has_load_barrier_nodes();
3247
3248 if (use->Opcode() == Op_LoadP && use->bottom_type()->isa_rawptr()) {
3249 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
3250 Node* loadp = use->fast_out(i);
3251 const Type* ut = loadp->bottom_type();
3252 if (loadp->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(loadp)) {
3253 if (has_load_barrier_nodes) {
3254 // Search for load barriers behind the load
3255 push_load_barrier(worklist, barrier_set, loadp);
3256 }
3257 worklist.push(loadp);
3258 }
3259 }
3260 }
3261 }
3262
3263 void PhaseCCP::push_load_barrier(Unique_Node_List& worklist, const BarrierSetC2* barrier_set, const Node* use) {
3264 add_users_to_worklist_if(worklist, use, [&](Node* u) {
3265 return barrier_set->is_gc_barrier_node(u);
3266 });
3267 }
3268
3269 // AndI/L::Value() optimizes patterns similar to (v << 2) & 3, or CON & 3 to zero if they are bitwise disjoint.
3270 // Add the AndI/L nodes back to the worklist to re-apply Value() in case the value is now a constant or shift
3271 // value changed.
3272 void PhaseCCP::push_and(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3273 const TypeInteger* parent_type = type(parent)->isa_integer(type(parent)->basic_type());
3274 uint use_op = use->Opcode();
3275 if (
3276 // Pattern: parent (now constant) -> (ConstraintCast | ConvI2L)* -> And
3277 (parent_type != nullptr && parent_type->is_con()) ||
3278 // Pattern: parent -> LShift (use) -> (ConstraintCast | ConvI2L)* -> And
3279 ((use_op == Op_LShiftI || use_op == Op_LShiftL) && use->in(2) == parent)) {
3280
3281 auto push_and_uses_to_worklist = [&](Node* n) {
3282 uint opc = n->Opcode();
3283 if (opc == Op_AndI || opc == Op_AndL) {
3284 push_if_not_bottom_type(worklist, n);
3285 }
3286 };
3287 auto is_boundary = [](Node* n) {
3288 return !(n->is_ConstraintCast() || n->Opcode() == Op_ConvI2L);
3289 };
3290 use->visit_uses(push_and_uses_to_worklist, is_boundary);
3291 }
3292 }
3293
3294 // CastII::Value() optimizes CmpI/If patterns if the right input of the CmpI has a constant type. If the CastII input is
3295 // the same node as the left input into the CmpI node, the type of the CastII node can be improved accordingly. Add the
3296 // CastII node back to the worklist to re-apply Value() to either not miss this optimization or to undo it because it
3297 // cannot be applied anymore. We could have optimized the type of the CastII before but now the type of the right input
3298 // of the CmpI (i.e. 'parent') is no longer constant. The type of the CastII must be widened in this case.
3299 void PhaseCCP::push_cast_ii(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3300 if (use->Opcode() == Op_CmpI && use->in(2) == parent) {
3301 Node* other_cmp_input = use->in(1);
3302 add_users_to_worklist_if(worklist, other_cmp_input, [&](Node* u) {
3303 return u->is_CastII() && not_bottom_type(u);
3304 });
3305 }
3306 }
3307
3308 void PhaseCCP::push_opaque_zero_trip_guard(Unique_Node_List& worklist, const Node* use) const {
3309 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
3310 push_if_not_bottom_type(worklist, use->unique_out());
3311 }
3312 }
3313
3314 //------------------------------do_transform-----------------------------------
3315 // Top level driver for the recursive transformer
3316 void PhaseCCP::do_transform() {
3317 // Correct leaves of new-space Nodes; they point to old-space.
3318 C->set_root( transform(C->root())->as_Root() );
3319 assert( C->top(), "missing TOP node" );
3320 assert( C->root(), "missing root" );
3321 }
3322
3323 //------------------------------transform--------------------------------------
3324 // Given a Node in old-space, clone him into new-space.
3325 // Convert any of his old-space children into new-space children.
3326 Node *PhaseCCP::transform( Node *n ) {
3327 assert(n->is_Root(), "traversal must start at root");
3328 assert(_root_and_safepoints.member(n), "root (n) must be in list");
3329
3330 ResourceMark rm;
3331 // Map: old node idx -> node after CCP (or nullptr if not yet transformed or useless).
3332 Node_List node_map;
3333 // Pre-allocate to avoid frequent realloc
3334 GrowableArray <Node *> transform_stack(C->live_nodes() >> 1);
3335 // track all visited nodes, so that we can remove the complement
3336 Unique_Node_List useful;
3337
3338 if (KillPathsReachableByDeadTypeNode) {
3339 for (uint i = 0; i < _maybe_top_type_nodes.size(); ++i) {
3340 Node* type_node = _maybe_top_type_nodes.at(i);
3341 if (type(type_node) == Type::TOP) {
3342 ResourceMark rm;
3343 type_node->as_Type()->make_paths_from_here_dead(this, nullptr, "ccp");
3344 }
3345 }
3346 } else {
3347 assert(_maybe_top_type_nodes.size() == 0, "we don't need type nodes");
3348 }
3349
3350 // Initialize the traversal.
3351 // This CCP pass may prove that no exit test for a loop ever succeeds (i.e. the loop is infinite). In that case,
3352 // the logic below doesn't follow any path from Root to the loop body: there's at least one such path but it's proven
3353 // never taken (its type is TOP). As a consequence the node on the exit path that's input to Root (let's call it n) is
3354 // replaced by the top node and the inputs of that node n are not enqueued for further processing. If CCP only works
3355 // through the graph from Root, this causes the loop body to never be processed here even when it's not dead (that
3356 // is reachable from Root following its uses). To prevent that issue, transform() starts walking the graph from Root
3357 // and all safepoints.
3358 for (uint i = 0; i < _root_and_safepoints.size(); ++i) {
3359 Node* nn = _root_and_safepoints.at(i);
3360 Node* new_node = node_map[nn->_idx];
3361 assert(new_node == nullptr, "");
3362 new_node = transform_once(nn); // Check for constant
3363 node_map.map(nn->_idx, new_node); // Flag as having been cloned
3364 transform_stack.push(new_node); // Process children of cloned node
3365 useful.push(new_node);
3366 }
3367
3368 while (transform_stack.is_nonempty()) {
3369 Node* clone = transform_stack.pop();
3370 uint cnt = clone->req();
3371 for( uint i = 0; i < cnt; i++ ) { // For all inputs do
3372 Node *input = clone->in(i);
3373 if( input != nullptr ) { // Ignore nulls
3374 Node *new_input = node_map[input->_idx]; // Check for cloned input node
3375 if( new_input == nullptr ) {
3376 new_input = transform_once(input); // Check for constant
3377 node_map.map( input->_idx, new_input );// Flag as having been cloned
3378 transform_stack.push(new_input); // Process children of cloned node
3379 useful.push(new_input);
3380 }
3381 assert( new_input == clone->in(i), "insanity check");
3382 }
3383 }
3384 }
3385
3386 // The above transformation might lead to subgraphs becoming unreachable from the
3387 // bottom while still being reachable from the top. As a result, nodes in that
3388 // subgraph are not transformed and their bottom types are not updated, leading to
3389 // an inconsistency between bottom_type() and type(). In rare cases, LoadNodes in
3390 // such a subgraph, might be re-enqueued for IGVN indefinitely by MemNode::Ideal_common
3391 // because their address type is inconsistent. Therefore, we aggressively remove
3392 // all useless nodes here even before PhaseIdealLoop::build_loop_late gets a chance
3393 // to remove them anyway.
3394 if (C->cached_top_node()) {
3395 useful.push(C->cached_top_node());
3396 }
3397 C->update_dead_node_list(useful);
3398 remove_useless_nodes(useful.member_set());
3399 _worklist.remove_useless_nodes(useful.member_set());
3400 C->disconnect_useless_nodes(useful, _worklist, &_root_and_safepoints);
3401
3402 Node* new_root = node_map[n->_idx];
3403 assert(new_root->is_Root(), "transformed root node must be a root node");
3404 return new_root;
3405 }
3406
3407 //------------------------------transform_once---------------------------------
3408 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
3409 Node *PhaseCCP::transform_once( Node *n ) {
3410 const Type *t = type(n);
3411 // Constant? Use constant Node instead
3412 if( t->singleton() ) {
3413 Node *nn = n; // Default is to return the original constant
3414 if( t == Type::TOP ) {
3415 // cache my top node on the Compile instance
3416 if( C->cached_top_node() == nullptr || C->cached_top_node()->in(0) == nullptr ) {
3417 C->set_cached_top_node(ConNode::make(Type::TOP));
3418 set_type(C->top(), Type::TOP);
3419 }
3420 nn = C->top();
3421 }
3422 if( !n->is_Con() ) {
3423 if( t != Type::TOP ) {
3424 nn = makecon(t); // ConNode::make(t);
3425 NOT_PRODUCT( inc_constants(); )
3426 } else if( n->is_Region() ) { // Unreachable region
3427 // Note: nn == C->top()
3428 n->set_req(0, nullptr); // Cut selfreference
3429 bool progress = true;
3430 uint max = n->outcnt();
3431 DUIterator i;
3432 while (progress) {
3433 progress = false;
3434 // Eagerly remove dead phis to avoid phis copies creation.
3435 for (i = n->outs(); n->has_out(i); i++) {
3436 Node* m = n->out(i);
3437 if (m->is_Phi()) {
3438 assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
3439 replace_node(m, nn);
3440 if (max != n->outcnt()) {
3441 progress = true;
3442 i = n->refresh_out_pos(i);
3443 max = n->outcnt();
3444 }
3445 }
3446 }
3447 }
3448 }
3449 replace_node(n,nn); // Update DefUse edges for new constant
3450 }
3451 return nn;
3452 }
3453
3454 // If x is a TypeNode, capture any more-precise type permanently into Node
3455 if (t != n->bottom_type()) {
3456 hash_delete(n); // changing bottom type may force a rehash
3457 n->raise_bottom_type(t);
3458 _worklist.push(n); // n re-enters the hash table via the worklist
3459 add_users_to_worklist(n); // if ideal or identity optimizations depend on the input type, users need to be notified
3460 }
3461
3462 // TEMPORARY fix to ensure that 2nd GVN pass eliminates null checks
3463 switch( n->Opcode() ) {
3464 case Op_CallStaticJava: // Give post-parse call devirtualization a chance
3465 case Op_CallDynamicJava:
3466 case Op_FastLock: // Revisit FastLocks for lock coarsening
3467 case Op_If:
3468 case Op_CountedLoopEnd:
3469 case Op_Region:
3470 case Op_Loop:
3471 case Op_CountedLoop:
3472 case Op_Conv2B:
3473 case Op_Opaque1:
3474 _worklist.push(n);
3475 break;
3476 default:
3477 break;
3478 }
3479
3480 return n;
3481 }
3482
3483 //---------------------------------saturate------------------------------------
3484 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
3485 const Type* limit_type) const {
3486 const Type* wide_type = new_type->widen(old_type, limit_type);
3487 if (wide_type != new_type) { // did we widen?
3488 // If so, we may have widened beyond the limit type. Clip it back down.
3489 new_type = wide_type->filter(limit_type);
3490 }
3491 return new_type;
3492 }
3493
3494 //------------------------------print_statistics-------------------------------
3495 #ifndef PRODUCT
3496 void PhaseCCP::print_statistics() {
3497 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants);
3498 }
3499 #endif
3500
3501
3502 //=============================================================================
3503 #ifndef PRODUCT
3504 uint PhasePeephole::_total_peepholes = 0;
3505 #endif
3506 //------------------------------PhasePeephole----------------------------------
3507 // Conditional Constant Propagation, ala Wegman & Zadeck
3508 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
3509 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
3510 NOT_PRODUCT( clear_peepholes(); )
3511 }
3512
3513 #ifndef PRODUCT
3514 //------------------------------~PhasePeephole---------------------------------
3515 PhasePeephole::~PhasePeephole() {
3516 _total_peepholes += count_peepholes();
3517 }
3518 #endif
3519
3520 //------------------------------transform--------------------------------------
3521 Node *PhasePeephole::transform( Node *n ) {
3522 ShouldNotCallThis();
3523 return nullptr;
3524 }
3525
3526 //------------------------------do_transform-----------------------------------
3527 void PhasePeephole::do_transform() {
3528 bool method_name_not_printed = true;
3529
3530 // Examine each basic block
3531 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
3532 Block* block = _cfg.get_block(block_number);
3533 bool block_not_printed = true;
3534
3535 for (bool progress = true; progress;) {
3536 progress = false;
3537 // block->end_idx() not valid after PhaseRegAlloc
3538 uint end_index = block->number_of_nodes();
3539 for( uint instruction_index = end_index - 1; instruction_index > 0; --instruction_index ) {
3540 Node *n = block->get_node(instruction_index);
3541 if( n->is_Mach() ) {
3542 MachNode *m = n->as_Mach();
3543 // check for peephole opportunities
3544 int result = m->peephole(block, instruction_index, &_cfg, _regalloc);
3545 if( result != -1 ) {
3546 #ifndef PRODUCT
3547 if( PrintOptoPeephole ) {
3548 // Print method, first time only
3549 if( C->method() && method_name_not_printed ) {
3550 C->method()->print_short_name(); tty->cr();
3551 method_name_not_printed = false;
3552 }
3553 // Print this block
3554 if( Verbose && block_not_printed) {
3555 tty->print_cr("in block");
3556 block->dump();
3557 block_not_printed = false;
3558 }
3559 // Print the peephole number
3560 tty->print_cr("peephole number: %d", result);
3561 }
3562 inc_peepholes();
3563 #endif
3564 // Set progress, start again
3565 progress = true;
3566 break;
3567 }
3568 }
3569 }
3570 }
3571 }
3572 }
3573
3574 //------------------------------print_statistics-------------------------------
3575 #ifndef PRODUCT
3576 void PhasePeephole::print_statistics() {
3577 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes);
3578 }
3579 #endif
3580
3581
3582 //=============================================================================
3583 //------------------------------set_req_X--------------------------------------
3584 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
3585 assert( is_not_dead(n), "can not use dead node");
3586 #ifdef ASSERT
3587 if (igvn->hash_find(this) == this) {
3588 tty->print_cr("Need to remove from hash before changing edges");
3589 this->dump(1);
3590 tty->print_cr("Set at i = %d", i);
3591 n->dump();
3592 assert(false, "Need to remove from hash before changing edges");
3593 }
3594 #endif
3595 Node *old = in(i);
3596 set_req(i, n);
3597
3598 // old goes dead?
3599 if( old ) {
3600 switch (old->outcnt()) {
3601 case 0:
3602 // Put into the worklist to kill later. We do not kill it now because the
3603 // recursive kill will delete the current node (this) if dead-loop exists
3604 if (!old->is_top())
3605 igvn->_worklist.push( old );
3606 break;
3607 case 1:
3608 if( old->is_Store() || old->has_special_unique_user() )
3609 igvn->add_users_to_worklist( old );
3610 break;
3611 case 2:
3612 if( old->is_Store() )
3613 igvn->add_users_to_worklist( old );
3614 if( old->Opcode() == Op_Region )
3615 igvn->_worklist.push(old);
3616 break;
3617 case 3:
3618 if( old->Opcode() == Op_Region ) {
3619 igvn->_worklist.push(old);
3620 igvn->add_users_to_worklist( old );
3621 }
3622 break;
3623 default:
3624 break;
3625 }
3626
3627 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old);
3628 }
3629 }
3630
3631 void Node::set_req_X(uint i, Node *n, PhaseGVN *gvn) {
3632 PhaseIterGVN* igvn = gvn->is_IterGVN();
3633 if (igvn == nullptr) {
3634 set_req(i, n);
3635 return;
3636 }
3637 set_req_X(i, n, igvn);
3638 }
3639
3640 //-------------------------------replace_by-----------------------------------
3641 // Using def-use info, replace one node for another. Follow the def-use info
3642 // to all users of the OLD node. Then make all uses point to the NEW node.
3643 void Node::replace_by(Node *new_node) {
3644 assert(!is_top(), "top node has no DU info");
3645 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
3646 Node* use = last_out(i);
3647 uint uses_found = 0;
3648 for (uint j = 0; j < use->len(); j++) {
3649 if (use->in(j) == this) {
3650 if (j < use->req())
3651 use->set_req(j, new_node);
3652 else use->set_prec(j, new_node);
3653 uses_found++;
3654 }
3655 }
3656 i -= uses_found; // we deleted 1 or more copies of this edge
3657 }
3658 }
3659
3660 //=============================================================================
3661 //-----------------------------------------------------------------------------
3662 void Type_Array::grow( uint i ) {
3663 assert(_a == Compile::current()->comp_arena(), "Should be allocated in comp_arena");
3664 if( !_max ) {
3665 _max = 1;
3666 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
3667 _types[0] = nullptr;
3668 }
3669 uint old = _max;
3670 _max = next_power_of_2(i);
3671 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
3672 memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
3673 }
3674
3675 //------------------------------dump-------------------------------------------
3676 #ifndef PRODUCT
3677 void Type_Array::dump() const {
3678 uint max = Size();
3679 for( uint i = 0; i < max; i++ ) {
3680 if( _types[i] != nullptr ) {
3681 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr();
3682 }
3683 }
3684 }
3685 #endif